Pith. sign in

REVIEW 6 major objections 5 minor 49 references

HetSSNet: Spatial-Spectral Heterogeneous Graph Learning Network for Panchromatic and Multispectral Images Fusion

T0 review · 6 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a heterogeneous graph encoding three pansharpening-specific relationships fuses PAN and multispectral images more accurately than CNN and Transformer models.

desk verdict Plausible new architecture, but the missing graph-to-image reconstruction step makes the central training loss undefined and the SOTA claims unreproducible as written. read the letter →

arxiv 2502.04623 v1 pith:QQY4KCU3 submitted 2025-02-07 cs.CV

Mengting Ma , Yizhen Jiang , Mengjiao Zhao , Jiaxin Li , Wei Zhang This is my paper · ORCID
classification cs.CV
keywords pansharpeningheterogeneousgraphconvolutionalnetworkspatial-spectralfusionremotesensingcontrastivelearningimage
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

HetSSNet argues that pansharpening—the fusion of a high-resolution panchromatic image with a low-resolution multispectral image—is better cast as learning on a heterogeneous graph than as convolution or attention on a regular pixel grid, because ground objects in satellite scenes are irregular. The paper proposes a network that builds a graph with three kinds of edges—spatial relations among panchromatic patches, spectral relations among multispectral bands, and cross relations between the two—and then learns a unified representation by combining all edge-type patterns locally and globally. Its central empirical claim is that this graph formulation outperforms CNN- and Transformer-based pansharpening methods on almost every metric on the WorldView-3, QuickBird, and GaoFen-2 datasets, including no-reference metrics on real full-resolution scenes. If correct, the result suggests that explicit relationship priors in non-Euclidean space are a stronger inductive bias for spatial-spectral fusion than uniform grid models.

What carries the argument

HetSS-Graph is an attributed multiplex heterogeneous graph: a graph with multiple node types and multiple edge types, where each node has an attached feature vector and the same pair of nodes can be connected by more than one kind of relation. Here there are two node types—PAN patch nodes and LR-MS band nodes—and three edge types built by k-nearest-neighbor search in a cosine-similarity feature space. The argument runs through the basic relationship pattern generation module, which uses XNOR and AND operations on the three adjacency matrices to produce up to seven distinct relationship patterns, and through the relationship pattern aggregation module, which fuses those patterns with weighted graph convolution at the local level and with a similarity-based graph convolution at the global level, supervised by a contrastive loss between the two views. This machinery lets the network combine multiple relationship patterns into one representation instead of treating a single edge type as sufficient.

What would settle it

Retrain the same HetSSNet on GaoFen-2 with the only change being that the k-nearest-neighbor edges are replaced by fixed spatial-neighborhood edges from the same patch grid; if PSNR and SAM do not get worse, the feature-similarity graph topology is not the cause of the reported gains.

Watch

Extended reading notes

Core claim

The paper's central contribution is the first spatial-spectral heterogeneous graph for pansharpening. It constructs an attributed multiplex heterogeneous graph whose nodes are panchromatic image patches and multispectral band patches, and whose edges express three pansharpening-specific priors: the spatial relationship within the PAN image, the intra-spectral relationship within the LR-MS image, and the spectral relationship between the two images. A basic relationship pattern generation module enumerates up to $2^3-1=7$ edge-type combinations by logically combining the three adjacency matrices, and a relationship pattern aggregation module combines these patterns from a local view (weighted sum of pattern matrices fed to a simplified graph convolution) and a global view (a similarity matrix built from pattern-count vectors, also fed to graph convolution), with a contrastive loss aligning the two views. The paper reports that this network achieves the best or second-best score on nearly every reduced-resolution metric across WorldView-3, QuickBird, and GaoFen-2, and the best no-reference scores on full-resolution scenes, and concludes that non-Euclidean heterogeneous graph learning is a viable and generalizing alternative to grid-based fusion.

Load-bearing premise

The method assumes that linking each image patch to its most similar patches in feature space captures the spatial and spectral relationships that fusion needs; if similar-looking patches are not the ones that should exchange information, the graph structure is arbitrary and the reported gains could come from extra model capacity rather than from that structure.

Editorial extensions

If this is right

  • If HetSSNet's gains are real, graph-based non-Euclidean processing becomes a credible third backbone for low-level image fusion, alongside CNN and Transformer.
  • The three-edge heterogeneous graph gives a reusable template for other multi-modal fusion problems where one modality supplies spatial detail and another supplies spectral or color information.
  • The exhaustive edge-type combination method implies that multiple relationship patterns, not just single edge types, can be jointly learned; this extends directly to other attributed multiplex graph tasks.
  • The local-global contrastive loss suggests that aligning two graph views of the same scene can improve reconstruction, a principle that transfers to other graph-based image restoration tasks.

Reading between the lines

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

  • A direct follow-up would replace the cosine-similarity k-nearest-neighbor edges with spatial-neighborhood edges or learned edges, keeping every other module fixed, to test whether the graph topology rather than the extra capacity is the source of the reported advantage.
  • The relationship-pattern enumeration grows combinatorially with the number of edge types; scaling to more spectral bands may require a learned or sampled subset of patterns, which the paper does not address.
  • A parallel testable hypothesis is that the heterogeneous graph advantage would be largest in scenes with many irregular objects and smallest in uniform scenes; the current aggregate metrics do not isolate that.
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

6 major / 5 minor

Summary. The manuscript proposes HetSSNet, a heterogeneous graph neural network for pansharpening, i.e., fusing a high-resolution panchromatic (PAN) image with a low-resolution multispectral (LR-MS) image to produce a high-resolution multispectral (HR-MS) image. The method constructs a heterogeneous graph whose nodes are patch-level features of PAN and band-wise features of LR-MS, with three edge types meant to encode spatial, intra-spectral, and spectral relationship priors. It then extracts up to seven basic relationship patterns via logical combinations of adjacency matrices, aggregates them with local and global graph convolutions, combines the two views with a contrastive loss, and trains with an L1 loss against the HR-MS ground truth. The paper claims state-of-the-art results on WorldView-3, QuickBird, and GaoFen-2 reduced-resolution benchmarks and on full-resolution generalization experiments.

Significance. If the reported results hold, HetSSNet would be a useful step toward graph-based modeling for low-level remote sensing fusion, and the explicit enumeration of pansharpening-specific relationship priors is a reasonable motivation. The paper is also honest in including ablation studies for the relationship-pattern generation module, the aggregation module, the number of layers, and the contrastive weight, which helps localize the source of the reported gains. However, the central empirical claim is currently undercut by a missing graph-to-image reconstruction step that makes the training loss undefined as written, by apparent data errors in Table 6, and by a comparison protocol that retrains baselines without their original training details. These issues must be resolved before the superiority claim can be accepted.

major comments (6)
  1. [Sec. 3.6, Eq. (10)] The training objective is not defined as written. Equation (10) states L1 = ||H − GT||_1, where Eq. (9) defines H as the final node representation in R^{n×d}, while GT is the target HR-MS image of size 128×128×4 per Table 1. The paper never specifies how the node-feature matrix is reshaped, projected, or reassembled into a four-band image, nor does it describe any decoder or patch-overlap aggregation. Because the nodes are described as unordered patch features in Sec. 3.3, this is not a minor shape mismatch: without an explicit graph-to-image mapping, the loss, the gradients, and all reported numbers in Tables 2, 3, and 6 are not reproducible from the manuscript. The authors must add the missing reconstruction head (or equivalent operation) and state its exact input/output dimensions.
  2. [Secs. 3.3–3.5, Eqs. (2)–(9)] There is a dimension inconsistency in the graph convolutions. The adjacency matrices are defined in Sec. 3.4 as A_r ∈ R^{5n×5n}, and U is the matrix of node features for V_P and V_L, which should have 5n rows (n PAN nodes and 4n LR-MS band nodes). Equations (2) and (7) then produce node representations of size 5n×d, but Eq. (9) states that the final H is in R^{n×d}. No reduction from 5n nodes to n nodes is described, and the notation n is used inconsistently (e.g., B ∈ R^{n×N} in Eq. (5) versus the 5n-dimensional adjacency matrices). This makes the network architecture, and in particular the graph-to-image step, even harder to interpret and reproduce.
  3. [Appendix D, Table 6] Table 6 contains exact duplicated rows across different datasets: the CTINN row has identical values (Dλ=0.072, Ds=0.114, QNR=0.834) for GaoFen-2 and WorldView-3, and the MSDDN row has identical values (0.149, 0.353, 0.710) for WorldView-3 and QuickBird. Since these are different sensors and scenes, such exact coincidences almost certainly indicate a data-handling or table-construction error. This undermines the full-resolution generalization claim that HetSSNet achieves the optimal outcomes for all indexes. The authors should correct these rows and recompute the affected comparisons.
  4. [Sec. 4.1, Benchmark] The comparison protocol is potentially unfair. The paper states that all comparison methods are re-trained on the adopted datasets without directly using the experimental details in the original articles. Retraining eleven learning-based baselines without their original training schedules, hyperparameters, and data splits can easily disadvantage them relative to HetSSNet, which is trained with a carefully tuned schedule (Sec. 4.3). The SOTA claim in Sec. 4.4 therefore requires either using official or author-provided results under a common protocol, or reporting the exact training configurations used for each baseline so that the comparison is verifiable.
  5. [Secs. 4.3–4.4 and Tables 2–6] No error bars, standard deviations, or significance tests are reported for any of the quantitative results, and the ablation tables appear to be based on single runs. Some reported margins are very small, e.g., QuickBird PSNR of 37.228 for HetSSNet versus 37.162 for BiMPan in Table 2. Without repeated runs or statistical testing, the claim of superiority in almost all metrics cannot be distinguished from training noise. The authors should report mean and standard deviation over at least three seeds, or otherwise justify that the margins are stable.
  6. [Sec. 3.3, Edge construction] The graph construction step defines the spatial relationship of the PAN image using k-nearest neighbors in cosine-similarity feature space, rather than spatial adjacency in the image plane. Since the paper motivates the graph by irregular ground objects, it should be demonstrated that this feature-similarity topology is actually better than a regular grid neighborhood or a spatially local graph. A simple ablation replacing the kNN graph with a spatial-neighborhood graph would make the contribution of the graph topology concrete. As it stands, the reported gains could in principle come from the additional network capacity rather than from the proposed heterogeneous graph structure.
minor comments (5)
  1. [Throughout] There are repeated typos and inconsistent spellings, including Transfromer for Transformer, Wordview-3 for WorldView-3, and non-European space for non-Euclidean space.
  2. [Sec. 3.3, Third edge type] The sentence describing the third type of edge is garbled: it says 'we add the third type of edge directed from v_i^b for all N(v_i^b) to v_i for all N(v_i)', without a clear specification of which source nodes connect to which target nodes. This should be rewritten as a precise set-builder definition.
  3. [Sec. 3.4] The XNOR/AND procedure for generating basic relationship patterns is described only in words; a small pseudocode block or a formal definition of the logical operations would help reproducibility, especially regarding how zero matrices are discarded and how the seven patterns are indexed.
  4. [Table 1 and Appendix B] The dataset statistics are inconsistent: Table 1 gives 35,725 training and 3,370 testing GaoFen-2 samples (39,095 total), while Appendix B states that the GaoFen-2 dataset contains 38,645 sets; the discrepancy should be explained or corrected.
  5. [Fig. 5 and Fig. 6 captions] The captions of the QuickBird and GaoFen-2 qualitative figures mention ARFNet, but ARFNet is not listed in the benchmarks of Sec. 4.1 and is only introduced in the full-resolution table; the authors should either add ARFNet to the benchmark description or remove it from the figures.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical SOTA claim rests on supervised training against held-out GT and unseen full-resolution scenes; the self-citation is a dataset reference and not load-bearing.

full rationale

The claimed SOTA result is produced by a standard supervised pipeline: the graph node features H are trained against the ground-truth HR-MS image through L1 = ||H − GT||1 (Eq. 10) and evaluated on held-out reduced-resolution test images and unseen full-resolution scenes (Tabs. 2 and 6). No trained parameter, learned adjacency weight, or loss term is itself the metric being reported, and the three relationship priors in Appendix B are empirical motivations (histogram correlations) for edge types, not fitted values that reappear as outputs. The only self-citation (Ma et al., 2024, for the datasets) is a data-source reference and is not load-bearing for the architecture or the numerical comparisons; all comparison methods were re-trained on the same adopted datasets. The main text does omit the graph-to-image reshaping/projection needed to compare H ∈ R^{n×d} with a 128×128×4 GT in Eq. (10), which is a reproducibility defect, but it is an under-specification rather than a circular reduction: it does not make the prediction identical to an input or to a fitted constant. Therefore no circular step is established.

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

The central claim rests on the choice of graph topology, the feature-space similarity criterion for edges, and several hyperparameters that are either omitted or selected by ablation on the test datasets. The network weights themselves are learned by supervised training, which is standard, but the missing implementation details and the absence of ablated k and d mean the reader cannot tell how much of the result depends on unstated choices.

free parameters (6)
  • k (k-NN neighbors) = not reported
    Number of neighbors used in graph construction (Section 3.3). Not specified or ablated.
  • gamma (contrastive loss weight) = 0.01
    Chosen by ablation on GaoFen-2 (Table 5), used in Eq. 11.
  • tau (contrastive temperature) = not reported
    Temperature in Eq. 8; not specified in the paper.
  • l (number of aggregation layers) = 2
    Chosen by ablation (Table 4).
  • feature dimension d = not reported
    Dimension of patch feature vectors in Section 3.3; not given.
  • patch size and overlap = not reported
    Patch division of PAN and LR-MS images (Section 3.3); not specified.
assumptions (3)
  • domain assumption The three edge types (PAN k-NN, LR-MS k-NN, cross edges) capture the needed spatial-spectral priors.
    Section 3.3 and Appendix B claim these priors are necessary; no comparison with alternate graph constructions is given.
  • domain assumption Histogram correlation via EMD distance is a valid measure of spatial-spectral relationship similarity.
    Appendix B uses this to justify Priors 1 and 2; the measure itself is not validated.
  • domain assumption Non-Euclidean graph structure is more suitable than grid structure for irregular ground objects.
    Motivates the entire approach, but is only supported by the final performance claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HetSSNet: Spatial-Spectral Heterogeneous Graph Learning Network for Panchromatic and Multispectral Images Fusion." pith.science (2026). https://pith.science/paper/QQY4KCU3

@misc{pith2026250204623,
  author       = {Pith},
  title        = {Pith review of: HetSSNet: Spatial-Spectral Heterogeneous Graph Learning Network for Panchromatic and Multispectral Images Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QQY4KCU3}},
  note         = {Machine review of arXiv:2502.04623}
}
read the original abstract

Remote sensing pansharpening aims to reconstruct spatial-spectral properties during the fusion of panchromatic (PAN) images and low-resolution multi-spectral (LR-MS) images, finally generating the high-resolution multi-spectral (HR-MS) images. In the mainstream modeling strategies, i.e., CNN and Transformer, the input images are treated as the equal-sized grid of pixels in the Euclidean space. They have limitations in facing remote sensing images with irregular ground objects. Graph is the more flexible structure, however, there are two major challenges when modeling spatial-spectral properties with graph: \emph{1) constructing the customized graph structure for spatial-spectral relationship priors}; \emph{2) learning the unified spatial-spectral representation through the graph}. To address these challenges, we propose the spatial-spectral heterogeneous graph learning network, named \textbf{HetSSNet}. Specifically, HetSSNet initially constructs the heterogeneous graph structure for pansharpening, which explicitly describes pansharpening-specific relationships. Subsequently, the basic relationship pattern generation module is designed to extract the multiple relationship patterns from the heterogeneous graph. Finally, relationship pattern aggregation module is exploited to collaboratively learn unified spatial-spectral representation across different relationships among nodes with adaptive importance learning from local and global perspectives. Extensive experiments demonstrate the significant superiority and generalization of HetSSNet.

Figures

Figures reproduced from arXiv: 2502.04623 by the authors.

Figure 1
Figure 1. The overview of HetSSNet. Our HetSSNet consists of three components: spatial-spectral heterogeneous graph construction, basic relationship pattern generation and relationship pattern aggregation. According to the provided relationship priors, we construct the spatial-spectral heterogeneous graph structure in non-Euclidean space. Based on the constructed graph, we generate a series of basic spatial-spectral relations… view at source ↗
Figure 2
Figure 2. Qualitative results of reduced-resolution scene on the WorldView-3 dataset. Top group: the fused results. Bottom group: the error between fused results and reference. role in node representation learning of the HetSSNet for spatial-spectral properties reconstruction. The number of aggregation layer. As shown in Tab. 4, firstly, the performance of HetSSNet increases with the increasing number of layers. When the numb… view at source ↗
Figure 3
Figure 3. The spatial distribution comparison between LR-MS/PAN and GT (target HR-MS) in terms of each spectral band. The band illustrated in the figure refers to the spectral band. The correlation coefficient represents the correlation between the two distribution curves, and the larger the value, the greater the correlation. Prior 2: In order to reconstruct spectral property for target HR-MS image, intra-spectra relationshi… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The distribution comparison between LR-MS’s spectral bands, and the pixel distribution comparison between GT’s spectral bands. The band illustrated in the figure denotes the spectral band. The correlation coefficient represents the correlation between the two distribut…
Figure 5
Figure 5. Figure 5: Qualitative results of reduced-resolution scene on the QuickBird dataset. Top group: the fused results. Bottom group: the error between fused results and reference. D. More quantitative and visualization results As shown in [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of reduced-resolution scene on the GaoFen-2 dataset. Top group: the fused results. Bottom group: the error between fused results and reference. as seen with BROVEY. Even though deep learning has brought some spatial detail improvements to pansharpen…
Figure 7
Figure 7. Figure 7: Visual comparison on the GaoFen-2 full-resolution scene, with some details magnified for better comparison. E. Complexity Analysis The time complexity of aggregating all basic spatial-spectral relationship patterns is O(Nn 2 ), and the time complexity of graph convolut…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 47 canonical work pages

  1. [1]

    Multispectral and panchromatic data fusion assessment without reference

    Alparone, L., Aiazzi, B., Baronti, S., Garzelli, A., Nencini, F., and Selva, M. Multispectral and panchromatic data fusion assessment without reference. Photogrammetric Engineering & Remote Sensing, 74 0 (2): 0 193--200, 2008

    work page 2008

  2. [2]

    and Anitha, J

    Asokan, A. and Anitha, J. Change detection techniques for remote sensing applications: a survey. Earth Science Informatics, pp.\ 143–160, Jun 2019

    work page 2019

  3. [3]

    Bandara, W. G. C. and Patel, V. M. Hypertransformer: A textural and spectral feature fusion transformer for pansharpening. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1767--1777, 2022

    work page 2022

  4. [4]

    Bandara, W. G. C., Valanarasu, J. M. J., and Patel, V. M. Spin road mapper: Extracting roads from aerial images via spatial and interaction space graph reasoning for autonomous driving. In 2022 International Conference on Robotics and Automation (ICRA), pp.\ 343--350. IEEE, 2022

    work page 2022

  5. [5]

    and Huang, B

    Cai, J. and Huang, B. Super-resolution-guided progressive pansharpening based on a deep convolutional neural network. IEEE Transactions on Geoscience and Remote Sensing, 59 0 (6): 0 5206--5220, 2020

    work page 2020

  6. [6]

    The use of intensity-hue-saturation transformations for merging spot panchromatic and multispectral image data

    Carper, W., Lillesand, T., Kiefer, R., et al. The use of intensity-hue-saturation transformations for merging spot panchromatic and multispectral image data. Photogrammetric Engineering and remote sensing, 56 0 (4): 0 459--467, 1990

    work page 1990

  7. [7]

    and Han, J

    Cheng, G. and Han, J. A survey on object detection in optical remote sensing images. ISPRS Journal of Photogrammetry and Remote Sensing, pp.\ 11–28, Jul 2016

    work page 2016

  8. [8]

    The fusion of panchromatic and multispectral remote sensing images via tensor-based sparse modeling and hyper-laplacian prior

    Deng, L.-J., Feng, M., and Tai, X.-C. The fusion of panchromatic and multispectral remote sensing images via tensor-based sparse modeling and hyper-laplacian prior. Information Fusion, 52: 0 76--89, 2019

    work page 2019

Show all 49 references
  1. [9]

    E., Scarpa, G., He, J., Zhang, Y., Chanussot, J., and Plaza, A

    Deng, L.-J., Vivone, G., Paoletti, M. E., Scarpa, G., He, J., Zhang, Y., Chanussot, J., and Plaza, A. Machine learning in pansharpening: A benchmark, from shallow to deep networks. IEEE Geoscience and Remote Sensing Magazine, 10 0 (3): 0 279--315, 2022

    2022

  2. [10]

    Content-adaptive non-local convolution for remote sensing pansharpening

    Duan, Y., Wu, X., Deng, H., and Deng, L.-J. Content-adaptive non-local convolution for remote sensing pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 27738--27747, 2024

    2024

  3. [11]

    MAGNN: metapath aggregated graph neural network for heterogeneous graph embedding

    Fu, X., Zhang, J., Meng, Z., and King, I. MAGNN: metapath aggregated graph neural network for heterogeneous graph embedding. In Huang, Y., King, I., Liu, T., and van Steen, M. (eds.), WWW '20: The Web Conference 2020, Taipei, Taiwan, April 20-24, 2020 , pp.\ 2331--2341. ACM / ...

    2020

  4. [12]

    Rain streak removal via dual graph convolutional network

    Fu, X., Qi, Q., Zha, Z.-J., Zhu, Y., and Ding, X. Rain streak removal via dual graph convolutional network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pp.\ 1352--1360, 2021

    2021

  5. [13]

    chromaticity

    Gillespie, A. R., Kahle, A. B., and Walker, R. E. Color enhancement of highly correlated images. ii. channel ratio and “chromaticity” transformation techniques. Remote Sensing of Environment, pp.\ 343–365, Aug 2003

    2003

  6. [14]

    Vision gnn: An image is worth graph of nodes

    Han, K., Wang, Y., Guo, J., Tang, Y., and Wu, E. Vision gnn: An image is worth graph of nodes. Advances in neural information processing systems, 35: 0 8291--8303, 2022

    2022

  7. [15]

    Gretnet: Gaussian retentive network for hyperspectral image classification

    Han, Z., Xu, S., Gao, L., Li, Z., and Zhang, B. Gretnet: Gaussian retentive network for hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 2024

    2024

  8. [16]

    Multi-scale dual-domain guidance network for pan-sharpening

    He, X., Yan, K., Zhang, J., Li, R., Xie, C., Zhou, M., and Hong, D. Multi-scale dual-domain guidance network for pan-sharpening. IEEE Transactions on Geoscience and Remote Sensing, 2023

    2023

  9. [17]

    Graph convolutional networks for hyperspectral image classification

    Hong, D., Gao, L., Yao, J., Zhang, B., Plaza, A., and Chanussot, J. Graph convolutional networks for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 59 0 (7): 0 5966--5978, 2020

    2020

  10. [18]

    Revisiting graph convolutional networks with mini-batch sampling for hyperspectral image classification

    Hong, D., Gao, L., Wu, X., Yao, J., and Zhang, B. Revisiting graph convolutional networks with mini-batch sampling for hyperspectral image classification. In 2021 11th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS), pp.\ 1--5, 2021

    2021

  11. [19]

    Bidomain modeling paradigm for pansharpening

    Hou, J., Cao, Q., Ran, R., Liu, C., Li, J., and Deng, L.-j. Bidomain modeling paradigm for pansharpening. In Proceedings of the 31st ACM International Conference on Multimedia, pp.\ 347--357, 2023

    2023

  12. [20]

    Graph relation network: Modeling relations between scenes for multilabel remote-sensing image classification and retrieval

    Kang, J., Fernandez-Beltran, R., Hong, D., Chanussot, J., and Plaza, A. Graph relation network: Modeling relations between scenes for multilabel remote-sensing image classification and retrieval. IEEE Transactions on Geoscience and Remote Sensing, 59 0 (5): 0 4355--4369, 2020

    2020

  13. [21]

    M., Chanussot, J., Condat, L., and Montanvert, A

    Khan, M. M., Chanussot, J., Condat, L., and Montanvert, A. Indusion: Fusion of multispectral and panchromatic images using the induction scaling technique. IEEE Geoscience and Remote Sensing Letters, 5 0 (1): 0 98--102, 2008

    2008

  14. [22]

    Deep learning in multimodal remote sensing data fusion: A comprehensive review

    Li, J., Hong, D., Gao, L., Yao, J., Zheng, K., Zhang, B., and Chanussot, J. Deep learning in multimodal remote sensing data fusion: A comprehensive review. International Journal of Applied Earth Observation and Geoinformation, 112: 0 102926, 2022

    2022

  15. [23]

    Local-global transformer enhanced unfolding network for pan-sharpening

    Li, M., Liu, Y., Xiao, T., Huang, Y., and Yang, G. Local-global transformer enhanced unfolding network for pan-sharpening. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI 2023, 19th-25th August 2023, Macao, SAR, China , pp.\...

    2023

  16. [24]

    Spatial pyramid based graph reasoning for semantic segmentation

    Li, X., Yang, Y., Zhao, Q., Shen, T., Lin, Z., and Liu, H. Spatial pyramid based graph reasoning for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 8950--8959, 2020

    2020

  17. [25]

    Liu, J. G. Smoothing filter-based intensity modulation: A spectral preserve image fusion technique for improving spatial details. International Journal of Remote Sensing, pp.\ 3461–3472, Jul 2002

    2002

  18. [26]

    Crocfun: Cross-modal conditional fusion network for pansharpening

    Ma, M., Hu, C., Zhang, H., Ma, X., Feng, T., and Zhang, W. Crocfun: Cross-modal conditional fusion network for pansharpening. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp.\ 3095--3099, 2024

    2024

  19. [27]

    3d graph neural networks for rgbd semantic segmentation

    Qi, X., Liao, R., Jia, J., Fidler, S., and Urtasun, R. 3d graph neural networks for rgbd semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pp.\ 5199--5208, 2017

    2017

  20. [28]

    Qin, A., Shang, Z., Tian, J., Wang, Y., Zhang, T., and Tang, Y. Y. Spectral--spatial graph convolutional networks for semisupervised hyperspectral image classification. IEEE Geoscience and Remote Sensing Letters, 16 0 (2): 0 241--245, 2018

    2018

  21. [29]

    S., Patil, S

    Sandhu, K. S., Patil, S. S., Pumphrey, M. O., and Carter, A. H. Multi-trait machine and deep learning models for genomic selection using spectral information in a wheat breeding program. Apr 2021

    2021

  22. [30]

    Image denoising with graph-convolutional neural networks

    Valsesia, D., Fracastoro, G., and Magli, E. Image denoising with graph-convolutional neural networks. In 2019 IEEE international conference on image processing (ICIP), pp.\ 2399--2403. IEEE, 2019

    2019

  23. [31]

    Hyperrealistic image inpainting with hypergraphs

    Wadhwa, G., Dhall, A., Murala, S., and Tariq, U. Hyperrealistic image inpainting with hypergraphs. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp.\ 3912--3921, 2021

    2021

  24. [32]

    Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images

    Wald, L., Ranchin, T., and Mangolini, M. Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images. Photogrammetric engineering and remote sensing, 63 0 (6): 0 691--699, 1997

    1997

  25. [33]

    Multiscale dynamic graph convolutional network for hyperspectral image classification

    Wan, S., Gong, C., Zhong, P., Du, B., Zhang, L., and Yang, J. Multiscale dynamic graph convolutional network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 58 0 (5): 0 3162--3177, 2019

    2019

  26. [34]

    Dynamic cross feature fusion for remote sensing pansharpening

    Wu, X., Huang, T.-Z., Deng, L.-J., and Zhang, T.-J. Dynamic cross feature fusion for remote sensing pansharpening. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), Mar 2022

    2021

  27. [35]

    Pansharpening via frequency-aware fusion network with explicit similarity constraints

    Xing, Y., Zhang, Y., He, H., Zhang, X., and Zhang, Y. Pansharpening via frequency-aware fusion network with explicit similarity constraints. IEEE Transactions on Geoscience and Remote Sensing, pp.\ 1–1, Jan 2023

    2023

  28. [36]

    Frequency-adaptive pan-sharpening with mixture of experts

    Xuanhua, H., Keyu, Y., Rui, L., Chengjun, X., Jie, Z., and Man, Z. Frequency-adaptive pan-sharpening with mixture of experts. In 2024 Conference on Innovative Applications of Artificial Intelligence (AAAI), 2024

    2024

  29. [37]

    When pansharpening meets graph convolution network and knowledge distillation

    Yan, K., Zhou, M., Liu, L., Xie, C., and Hong, D. When pansharpening meets graph convolution network and knowledge distillation. IEEE Transactions on Geoscience and Remote Sensing, 60: 0 1--15, 2022

    2022

  30. [38]

    Memory-augmented deep conditional unfolding network for pan-sharpening

    Yang, G., Zhou, M., Yan, K., Liu, A., Fu, X., and Wang, F. Memory-augmented deep conditional unfolding network for pan-sharpening. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1788--1797, 2022

    2022

  31. [39]

    Panflownet: A flow-based deep network for pan-sharpening

    Yang, G., Cao, X., Xiao, W., Zhou, M., Liu, A., chen, X., and Meng, D. Panflownet: A flow-based deep network for pan-sharpening. May 2023

    2023

  32. [40]

    Pannet: A deep network architecture for pan-sharpening

    Yang, J., Fu, X., Hu, Y., Huang, Y., Ding, X., and Paisley, J. Pannet: A deep network architecture for pan-sharpening. In Proceedings of the IEEE international conference on computer vision, pp.\ 5449--5457, 2017

    2017

  33. [41]

    Multiplex heterogeneous graph convolutional network

    Yu, P., Fu, C., Yu, Y., Huang, C., Zhao, Z., and Dong, J. Multiplex heterogeneous graph convolutional network. In Zhang, A. and Rangwala, H. (eds.), KDD '22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022 , pp.\...

    2022

  34. [42]

    Multiplex heterogeneous graph convolutional network

    Yu, P., Fu, C., Yu, Y., Huang, C., Zhao, Z., and Dong, J. Multiplex heterogeneous graph convolutional network. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 2377--2387, 2022 b

    2022

  35. [43]

    Zhang, L., Li, X., Arnab, A., Yang, K., Tong, Y., and Torr, P. H. Dual graph convolutional network for semantic segmentation. arXiv preprint arXiv:1909.06121, 2019

    1909 arXiv

  36. [44]

    Zhao, L., Peng, X., Tian, Y., Kapadia, M., and Metaxas, D. N. Semantic graph convolutional networks for 3d human pose regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3425--3435, 2019

    2019

  37. [45]

    Panformer: A transformer based model for pan-sharpening

    Zhou, H., Liu, Q., and Wang, Y. Panformer: A transformer based model for pan-sharpening. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pp.\ 1--6, 2022 a

    2022

  38. [46]

    L., and Silander, J

    Zhou, J., Civco, D. L., and Silander, J. A. A wavelet transform method to merge landsat tm and spot panchromatic data. International Journal of Remote Sensing, pp.\ 743–757, Jan 1998

    1998

  39. [47]

    Pan-sharpening with customized transformer and invertible neural network

    Zhou, M., Huang, J., Fang, Y., Fu, X., and Liu, A. Pan-sharpening with customized transformer and invertible neural network. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pp.\ 3553--3561, 2022 b

    2022

  40. [48]

    Spatial-frequency domain information integration for pan-sharpening

    Zhou, M., Huang, J., Yan, K., Yu, H., Fu, X., Liu, A., Wei, X., and Zhao, F. Spatial-frequency domain information integration for pan-sharpening. In European Conference on Computer Vision, pp.\ 274--291, 2022 c

    2022

  41. [49]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.