SWIFT: A General Sensitive Weight Identification Framework for Fast Sensor-Transfer Pansharpening

Chenxi Sun; Haoyu Chen; Liang-Jian Deng; Tianyu Xin; Yubo Zeng; Zeyu Xia

REVIEW 3 major objections 5 minor 1 cited by

SWIFT: A General Sensitive Weight Identification Framework for Fast Sensor-Transfer Pansharpening

T0 review · 3 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 pre-trained pansharpening model can be re-targeted to a new satellite sensor in about one minute by updating only the 30% of weights most sensitive to the domain shift, using just 3% of target images.

desk verdict A sensible two-step adaptation framework with strong reported gains, but the paper never defines the loss used to compute the gradients that drive the core sensitivity score, which is a load-bearing gap. read the letter →

arxiv 2507.20311 v1 pith:WGRPQDPF submitted 2025-07-27 cs.CV

Zeyu Xia , Chenxi Sun , Tianyu Xin , Yubo Zeng , Haoyu Chen , Liang-Jian Deng This is my paper · ORCID

classification cs.CV

keywords pansharpeningcross-sensoradaptationtransferlearningparameter-efficientfine-tuninggradientsensitivitysampleselectionremotesensing

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper takes on a practical bottleneck in pansharpening: models trained on one satellite's images lose accuracy when pointed at another sensor, and fixing that by full retraining costs hours. SWIFT is a plug-and-play procedure that first picks a 3% subset of target image pairs using a density-aware farthest-point sampling, then uses gradients from that subset to score every weight's sensitivity to the domain shift and updates only the most sensitive roughly 30% of weights. Across multiple models and sensor pairs, the adapted models reach or exceed full-retraining quality in about one minute on a single GPU. The reason this matters is that it makes cross-sensor deployment a weight-only update rather than a redesign or a long training run.

What carries the argument

The load-bearing object is the composite sensitivity score $S(\theta_j) = \alpha\cdot\mathrm{MAG}(\theta_j) + \beta\cdot(1-\mathrm{STD}(\theta_j)) + \gamma\cdot\mathrm{GDC}(\theta_j)$, computed over $M$ microbatches of the selected subset. MAG is the mean absolute expected gradient, GDC measures how consistently positive-versus-negative update signs appear across samples, and STD is the gradient's standard deviation, penalized because stable update intentions are preferred. Ranking parameters by this score and applying a sharpness-based dynamic selection ratio yields the subset $\theta_{\mathrm{SWIFT}}$ that gets updated. The companion machinery is Density-Aware Farthest Point Sampling, which weights each candidate's distance to already-selected samples by a density term so the 3% subset covers rare and typical target structures alike.

What would settle it

Build the 3% subset from a single land-cover class of a target sensor not used in the paper; if adapted models then fail to beat direct transfer on class-balanced held-out images, the density-aware coverage assumption is the load-bearing component.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-sensor pansharpening adaptation can be reduced to a targeted weight update guided by gradient statistics on a tiny sample. Given a source-pretrained model and unlabeled target pairs, SWIFT selects 3% of target data by balancing sparse-region uniqueness with dense-region representativeness, then computes for each parameter a sensitivity score that combines average gradient magnitude, sign-consistency across microbatches, and inverse gradient variance, plus a dynamic threshold based on the sharpness of the magnitude distribution. The framework keeps the architecture fixed and retrains only the selected weight subset on the small sample. The experiments report that this one-minute procedure matches or beats full retraining on WorldView-2 and QuickBird targets for models including PanNet, FusionNet, U2Net, SSDiff, ADWM, and WFANet.

Load-bearing premise

The framework assumes that gradient statistics computed from a 3% sample of the target domain correctly rank which weights matter for the entire target domain.

Editorial extensions

If this is right

If SWIFT holds, routine cross-sensor deployment becomes a one-minute weight-only update instead of a multi-hour retraining run.
Existing deployed architectures need no change, so already-installed models can be updated without a redeployment cycle.
The cost savings grow with model size: for SSDiff the paper reports adaptation falling from close to 10 hours to under 20 minutes.
Because only about 30% of weights are touched, the rest of the network can stay frozen, lowering GPU memory and update risk.

Reading between the lines

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

The same two-step pattern, coverage-aware sample selection followed by gradient-based parameter scoring, could generalize to other low-level vision tasks such as super-resolution or denoising where a model trained on one camera or degradation moves to another.
The fact that roughly 30% of weights suffice hints that sensor shift lives in a low-dimensional subspace of parameter space; if true, sensitivity scores computed on 3% data could also guide continual-learning updates that avoid catastrophic forgetting.
A practical extension would be to use the sensitivity ranking not just to pick weights but to order them, so that under tighter time budgets one could adapt with even fewer tunable parameters and read off the quality-versus-budget trade-off.

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.

Desk Editor's Note

read the letter

The core idea is useful and the empirical story is strong: select 3% of target-domain samples with a density-aware FPS variant, then use gradient statistics to pick and update roughly 30% of parameters, leaving the architecture untouched. Reported results show large gains over direct transfer and often match or beat full retraining, at a fraction of the adaptation cost. That is a practical contribution for cross-sensor pansharpening, and the time/parameter budgets are reported as marginal adaptation cost, which is honest.

What is actually new is the combination, not the ingredients. DA-FPS reduces to standard FPS at alpha=0, and gradient-based importance scoring is straight out of the pruning/OB D tradition. But the joint pipeline for sensor transfer is not in the cited literature, and the cross-sensor experiments with 3% data are new.

Now the soft spots, in proportion. The biggest one is the stress-test concern, and I think it lands hard: the method section defines the target domain as unlabeled pairs T = {Y_T, P_T} and never specifies the objective function whose gradients are computed in Eq. 7. A forward/backward pass requires a loss. If they use Wald-simulated HRMS references, the data are not unlabeled; if they use a self-supervised PAN/LRMS consistency loss, it must be stated because different losses will rank parameters differently. Without the loss, Eq. 7 is undefined and the entire adaptation mechanism cannot be reproduced or checked. This is the load-bearing step, not a cosmetic omission.

The secondary issues the reader flagged are real but addressable: "Figure N" and "Table ??" placeholders, a garbled Table 5, missing PanAdapter baseline despite discussing it, no code or data release, and hyperparameters (sigma, alpha, beta, gamma, eta_min, eta_max, M) mostly unspecified. The abstract's "approximately one minute" is also an overstatement for SSDiff (about 18 minutes) and WFANet (about 27 minutes), though the per-image comparison is still favorable.

Who is this for? People working on efficient adaptation in low-level vision and remote sensing. The idea is worth engaging, and the results are large enough to justify referee time. But the missing loss definition is a major revision, not a polish item. I would send it to reviewers with a clear request to fix that, along with the placeholders and the hyperparameter table.

Recommendation: engage, but the authors must specify the loss and release at least the selection/adaptation code before this can be credited as a reliable method.

Referee Report

3 major / 5 minor

Summary. The paper proposes SWIFT, a two-stage framework for fast cross-sensor pansharpening adaptation. In the first stage, Density-Aware Farthest Point Sampling (DA-FPS) selects 3% of the target-domain samples by balancing density and coverage on the data manifold. In the second stage, the framework computes per-parameter gradient statistics (magnitude, direction consistency, standard deviation) on microbatches of that subset, ranks parameters by a composite sensitivity score, and updates only the most sensitive ~30% of parameters. Experiments on QuickBird, GaoFen-2, WorldView-2, and WorldView-3 report adaptation times of roughly one to twenty minutes, with HQNR improvements over full retraining for several model/dataset pairs (e.g., our.FusionNet reaches HQNR 0.934 vs 0.840 on QB-GF2).

Significance. If the mechanism is well-defined, SWIFT is a practically valuable contribution: it is model-agnostic, preserves the model architecture, and the time comparisons are presented as marginal adaptation cost rather than total training cost. The gains are substantial for several baselines, with e.g. our.FusionNet improving HQNR from 0.840 to 0.934 on QB-GF2 and our.WFANet reducing ERGAS by about 21 on the reduced-resolution QB task. The central claim, however, depends entirely on the gradient-based sensitivity score being computed from a well-defined objective, and the paper does not currently specify that objective.

major comments (3)

[Sensitive Parameter Selection and Efficient Model Adaptation, Eq. (7)] The sensitivity score S(θj) is computed from gradients Gj obtained by "a full forward and backward pass" on microbatches of the selected target subset, but the target domain is defined as unlabeled image pairs T = {YT, PT} and no loss function is ever specified. Pansharpening supervision normally requires an HRMS reference; without one, the gradients used in Eqs. (4)-(7) are undefined. If the authors instead use Wald-simulated target HRMS references or a self-supervised loss involving PAN/LRMS consistency, that loss must be stated explicitly, because different losses will rank parameters differently and every result in Tables 2-4 depends on this ranking.
[Tables 2-4 and Generalization Ability, Training Details and Benchmark] The abstract claims SWIFT "substantially outperform[s] direct-transfer baselines," but the baseline rows in Tables 2-4 are described in the text as full retraining on the target domain ("for baseline models ... it is the full retraining time on the target domain"), not as direct transfer. The comparison is also not uniformly favorable: in Table 3, our.PanNet on GF2-QB has worse SCC (0.655 vs 0.757) and Q2N (0.315 vs 0.402) than the baseline. The authors should state precisely what each baseline row represents and qualify the "comparable to or better than full retraining" claim accordingly.
[Ablation Study, Dynamic Parameter Selection] The key ablation for dynamic parameter selection is not verifiable: the text says the results are "shown in Figure N" and that the dynamic method retains only 34.1% of parameters, but the figure is a placeholder. The surrounding text also says "Figure N presents the ablation results" without a real reference, so the claim that dynamic selection surpasses every fixed-ratio selection and matches full fine-tuning cannot be checked. In addition, Table 5's "Random" row has misaligned entries (a value of 0.747 appears in the Dλ column), and the caption refers to both reduced-resolution and full-resolution samples without indicating which resolution the columns report.

minor comments (5)

[Eq. (1)] The KDE density computation in Eq. (1) is O(N^2) in the number of target samples; for N=10,000 patches the paper should report the wall-clock cost of DA-FPS separately so that the "approximately one minute" adaptation claim can be assessed.
[Eqs. (3) and (7)] The symbol α is used both as the density-balance factor in Eq. (3) and as the sensitivity weight in Eq. (7); these are different hyperparameters and should be renamed to avoid confusion.
[Eq. (8)] Eq. (8) defines H as std({MAGj}) + (max({MAGj}) - median({MAGj})) and calls the second term "skewness"; this is not the standard definition of skewness and should be renamed or justified.
[References] The text uses "FusionNet" for two different works: Quan et al. 2021 (connectomics segmentation) in the introduction and Deng et al. 2021 (pansharpening) in the related work and benchmark; the benchmark should explicitly identify which FusionNet is used. The SSDiff reference also appears twice (Zhong et al. 2024 and 2025) with nearly identical titles and should be consolidated.
[Main Experimental Results] The sentence "presented in Tables 2, 3, 4, ??" contains a literal placeholder that should be replaced with the actual table number.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: sensitivity scores are heuristic gradient statistics, and all headline results are measured on held-out test images.

full rationale

The paper's derivation chain is not circular. The sensitivity score S(θj) (Eq. 7) is a composite of gradient magnitude, direction consistency, and standard deviation (Eqs. 4-6); it is not defined in terms of, nor optimized against, the reported quality metrics (HQNR, SAM, ERGAS, SCC, Q2N). The parameter subset θ_SWIFT is selected by ranking this gradient-based score and by the dynamic sharpness rule in Eqs. 8-10, neither of which embeds the target-domain test result as an input. The data subset t is chosen by density-weighted farthest point sampling (Eqs. 1-3), a distributional heuristic independent of downstream fusion quality. Evaluation in Tables 2-4 is on separate 20-image reduced- and full-resolution test sets, so the reported performance comparisons are externally grounded rather than being forced by the fitting procedure. The paper does cite several works by overlapping authors (FusionNet, U2Net, ADWM, SSDiff, PanAdapter), but these are baseline models and prior architecture works, not load-bearing justifications or uniqueness theorems used to mandate SWIFT's design. The main genuine weakness is that the method section never states the loss whose gradients feed Eq. 7, which is a reproducibility and correctness risk, but an underspecified loss is not a circular reduction. Likewise, the absence of a described validation split is an overfitting risk, not a circularity. No reported quantity is equal to an input by construction.

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

The framework rests on the assumption that a 3% density-aware sample captures the target manifold and that gradient statistics on that subset reveal which weights to update. The heuristic composite score (Eq. 7) and dynamic thresholding (Eqs. 8-9) introduce several unspecified hyperparameters, but no new physical entities.

free parameters (7)

sigma (KDE bandwidth)
Controls kernel bandwidth in density estimate Eq. (1); value not reported.
alpha (density balance in Eq. 3)
Balances coverage vs density in DA-FPS; value not reported. Symbol is reused later for sensitivity weight.
beta (sensitivity weight in Eq. 7)
Weight for gradient stability term in composite sensitivity score; value not reported.
gamma (sensitivity weight in Eq. 7)
Weight for gradient direction consistency; value not reported.
r (sampling ratio) = 3% (claimed)
Target-domain sample fraction; stated as 3% in abstract and experiments.
eta_min, eta_max (dynamic selection ratio bounds)
Bounds for parameter selection ratio in Eq. (9); values not reported.
M (number of microbatches)
Microbatch count used to compute gradient statistics; not reported.

assumptions (4)

domain assumption Kernel Density Estimation and Farthest Point Sampling (Eqs. 1-3) provide a faithful low-dimensional manifold description of the target domain from a handful of samples.
The method assumes that the 3% subset preserves the essential target distribution; if the manifold structure is misestimated, both data selection and gradient statistics are biased.
ad hoc to paper Gradient magnitude, direction consistency, and variance computed over microbatches indicate parameter importance for the domain shift.
Eqs. (4)-(6) define MAG, GDC, STD; these are heuristic measures inspired by optimal brain damage but not derived from any theory of transfer.
domain assumption Updating ~30% of parameters while freezing the rest is sufficient to adapt the model, i.e., the domain shift is low-dimensional in weight space.
This is the core premise of the partial fine-tuning strategy (Section 'Sensitive Parameter Selection and Efficient Model Adaptation').
standard math Standard pansharpening quality metrics (SAM, ERGAS, SCC, Q2N/Q8, QNR) are reliable for comparing fusion quality.
Used throughout the evaluation; no new metric is introduced.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SWIFT: A General Sensitive Weight Identification Framework for Fast Sensor-Transfer Pansharpening." pith.science (2026). https://pith.science/paper/WGRPQDPF

@misc{pith2026250720311,
  author       = {Pith},
  title        = {Pith review of: SWIFT: A General Sensitive Weight Identification Framework for Fast Sensor-Transfer Pansharpening},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WGRPQDPF}},
  note         = {Machine review of arXiv:2507.20311}
}

read the original abstract

Pansharpening aims to fuse high-resolution panchromatic (PAN) images with low-resolution multispectral (LRMS) images to generate high-resolution multispectral (HRMS) images. Although deep learning-based methods have achieved promising performance, they generally suffer from severe performance degradation when applied to data from unseen sensors. Adapting these models through full-scale retraining or designing more complex architectures is often prohibitively expensive and impractical for real-world deployment. To address this critical challenge, we propose a fast and general-purpose framework for cross-sensor adaptation, SWIFT (Sensitive Weight Identification for Fast Transfer). Specifically, SWIFT employs an unsupervised sampling strategy based on data manifold structures to balance sample selection while mitigating the bias of traditional Farthest Point Sampling, efficiently selecting only 3\% of the most informative samples from the target domain. This subset is then used to probe a source-domain pre-trained model by analyzing the gradient behavior of its parameters, allowing for the quick identification and subsequent update of only the weight subset most sensitive to the domain shift. As a plug-and-play framework, SWIFT can be applied to various existing pansharpening models. Extensive experiments demonstrate that SWIFT reduces the adaptation time from hours to approximately one minute on a single NVIDIA RTX 4090 GPU. The adapted models not only substantially outperform direct-transfer baselines but also achieve performance competitive with, and in some cases superior to, full retraining, establishing a new state-of-the-art on cross-sensor pansharpening tasks for the WorldView-2 and QuickBird datasets.

Figures

Figures reproduced from arXiv: 2507.20311 by the authors.

**Figure 1.** The SWIFT framework and its performance impact. Top: The two-step strategy of SWIFT. Bottom: HQNR comparison of representative models with and without SWIFT enhancement. cess of fusing these two kinds of images to generate highresolution multispectral (HRMS) images is known as pansharpening. Over the past few decades, pansharpening methods have evolved considerably, transitioning from traditional approaches to m… view at source ↗

**Figure 2.** Overall Framework of the SWIFT framework [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Visual Fusion image and Error maps on QB dataset (reduced data). For the error maps, blue indicate low error. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Comparison of Weighted MMD (Maximum Mean [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Performance and efficiency comparison of our dy [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

DMAConv: Dual Mask-Adaptive Convolution for Remote Sensing Pansharpening
cs.CV 2025-12 conditional novelty 5.0 of 10

A dual-branch mask-adaptive convolution (Bi2MAC/DMAConv) reduces pansharpening cost by assigning redundant pixels to a global kernel and heterogeneous pixels to pixel-wise kernels, claiming efficiency and SOTA gains.

Reference graph

Works this paper leans on

32 extracted references · 30 canonical work pages · cited by 1 Pith paper

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Y.; Nam, M.-R.; and Kim, H

Choi, M.; Kim, R. Y.; Nam, M.-R.; and Kim, H. O. 2005. Fusion of multispectral and panchromatic Satellite images using the curvelet transform. IEEE Geoscience and Remote Sensing Letters, 2(2): 136--140

work page 2005
[4]

W.; Graham, M.; Poorthuis, A.; Shelton, T.; Stephens, M.; Wilson, M

Crampton, J. W.; Graham, M.; Poorthuis, A.; Shelton, T.; Stephens, M.; Wilson, M. W.; and Zook, M. 2013. Beyond the geotag: Situating ‘big data’ and leveraging the potential of the geoweb. Cartography and Geographic Information Science, 40(2): 130--139

work page 2013
[5]

L.; Denker, J

Cun, Y. L.; Denker, J. S.; and Solla, S. A. 1989. Optimal Brain Damage. In NIPS'89: Proceedings of the 3rd International Conference on Neural Information Processing Systems, 598--605. Association for Computing Machinery

work page 1989
[6]

Deng, L.-J.; Vivone, G.; Jin, C.; and Chanussot, J. 2021. Detail Injection-Based Deep Convolutional Neural Networks for Pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 59(8): 6995--7010

work page 2021
[7]

Fitzner, D.; Sester, M.; Haberlandt, U.; and Rabiei, E. 2013. Rainfall estimation with a geosensor network of cars: Theoretical considerations and first results. Photogrammetrie, Fernerkundung, Geoinformation, 2013(2): 93--103

work page 2013
[8]

Garzelli, A.; and Nencini, F. 2009. Hypercomplex Quality Assessment of Multi/Hyperspectral Images. IEEE Geoscience and Remote Sensing Letters, 6(4): 662--665

work page 2009

Show all 32 references

[9]

He, Y.; Zhong, Y.; and Ma, A. 2019. A deep spatial-spectral network for pansharpening. Remote Sensing, 11(17): 2061

2019
[10]

Huang, J.; Chen, H.; Ren, J.; Peng, S.; and Deng, L. 2025 a . A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening. In Proceedings of the Computer Vision and Pattern Recognition Conference, 7447--7456

2025
[11]

Huang, J.; Huang, R.; Xu, J.; Pen, S.; Duan, Y.; and Deng, L. 2025 b . Wavelet-Assisted Multi-Frequency Attention Network for Pansharpening. arXiv:2502.04903

2025 arXiv
[12]

Li, Z.; Chen, H.; Li, J.; et al. 2024. FusFormer: Global and detail feature fusion transformer for semantic segmentation of small objects. Multimedia Tools and Applications, 83: 88717--88744

2024
[13]

Masi, G.; Cozzolino, D.; Verdoliva, L.; and Scarpa, G. 2016. Pansharpening by Convolutional Neural Networks. Remote Sensing, 8(7)

2016
[14]

Meng, Q.; Shi, W.; Li, S.; and Zhang, L. 2023. PanDiff: A Novel Pansharpening Method Based on Denoising Diffusion Probabilistic Model. IEEE Transactions on Geoscience and Remote Sensing, 61: 1--17

2023
[15]

Otazu, X.; Gonzalez-Audicana, M.; Fors, O.; and Nunez, J. 2005. Introduction of sensor spectral response into image fusion methods. Application to wavelet-based methods. IEEE Transactions on Geoscience and Remote Sensing, 43(10): 2376--2385

2005
[16]

Peng, S.; Guo, C.; Wu, X.; and Deng, L.-J. 2023. U2Net: A General Framework with Spatial-Spectral-Integrated Double U-Net for Image Fusion. In Proceedings of the 31st ACM International Conference on Multimedia, MM '23, 3219–3227. New York, NY, USA: Association for Computing Ma...

2023
[17]

M.; Hildebrand, D

Quan, T. M.; Hildebrand, D. G. C.; and Jeong, W.-K. 2021. Fusionnet: A deep fully residual convolutional neural network for image segmentation in connectomics. Frontiers in Computer Science, 3: 613981

2021
[18]

Rui, X.; Cao, X.; Pang, L.; Zhu, Z.; Yue, Z.; and Meng, D. 2024. Unsupervised hyperspectral pansharpening via low-rank diffusion model. Information Fusion, 107: 102325

2024
[19]

Tian, X.; Chen, Y.; Yang, C.; and Ma, J. 2022. Variational Pansharpening by Exploiting Cartoon-Texture Similarities. IEEE Transactions on Geoscience and Remote Sensing, 60: 1--16

2022
[20]

A.; Restaino, R.; and Wald, L

Vivone, G.; Alparone, L.; Chanussot, J.; Dalla Mura, M.; Garzelli, A.; Licciardi, G. A.; Restaino, R.; and Wald, L. 2015 a . A Critical Comparison Among Pansharpening Algorithms. IEEE Transactions on Geoscience and Remote Sensing, 53(5): 2565--2586

2015
[21]

A.; Restaino, R.; and Wald, L

Vivone, G.; Alparone, L.; Chanussot, J.; Dalla Mura, M.; Garzelli, A.; Licciardi, G. A.; Restaino, R.; and Wald, L. 2015 b . A Critical Comparison Among Pansharpening Algorithms. IEEE Transactions on Geoscience and Remote Sensing, 53(5): 2565--2586

2015
[22]

Vivone, G.; Restaino, R.; and Chanussot, J. 2018. A Regression-Based High-Pass Modulation Pansharpening Approach. IEEE Transactions on Geoscience and Remote Sensing, 56(2): 984--996

2018
[23]

Wald, L. 2002. Data Fusion: Definitions and Architectures : Fusion of Images of Different Spatial Resolutions. Presses de l' \'E cole des Mines. ISBN 9782911762383

2002
[24]

Wu, R.; Zhang, Z.; Deng, S.; Duan, Y.; and Deng, L.-J. 2025. Panadapter: Two-stage fine-tuning with spatial-spectral priors injecting for pansharpening. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 8450--8459

2025
[25]

Yang, G.; Cao, X.; Xiao, W.; Zhou, M.; Liu, A.; Chen, X.; and Meng, D. 2023. PanFlowNet: A Flow-Based Deep Network for Pan-Sharpening. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 16857--16867

2023
[26]

Yang, J.; Fu, X.; Hu, Y.; Huang, Y.; Ding, X.; and Paisley, J. 2017. PanNet: A Deep Network Architecture for Pan-Sharpening. In 2017 IEEE International Conference on Computer Vision (ICCV), 1753--1761

2017
[27]

H.; Goetz, A

Yuhas, R. H.; Goetz, A. F. H.; and Boardman, J. W. 1992. Discrimination among semi-arid landscape endmembers using the Spectral Angle Mapper (SAM) algorithm. In Summaries of the Third Annual JPL Airborne Geoscience Workshop. Volume 1: AVIRIS Workshop

1992
[28]

Zhang, H.; and Ma, J. 2021. GTP-PNet: A residual learning network based on gradient transformation prior for pansharpening. ISPRS Journal of Photogrammetry and Remote Sensing, 172: 223--239

2021
[29]

Zhong, Y.; Wu, X.; Cao, Z.; Dou, H.-X.; and Deng, L.-J. 2024. Ssdiff: Spatial-spectral integrated diffusion model for remote sensing pansharpening. Advances in Neural Information Processing Systems, 37: 77962--77986

2024
[30]

Zhong, Y.; Wu, X.; Cao, Z.; Dou, H.-X.; and Deng, L.-J. 2025. SSDiff: spatial-spectral integrated diffusion model for remote sensing pansharpening. In Proceedings of the 38th International Conference on Neural Information Processing Systems, NIPS '24. Red Hook, NY, USA: Curran...

2025
[31]

Zhou, H.; Liu, Q.; and Wang, Y. 2022. PanFormer: A Transformer Based Model for Pan-Sharpening. In 2022 IEEE International Conference on Multimedia and Expo (ICME), 1--6

2022
[32]

T.; Civco, D

Zhou, J. T.; Civco, D. L.; and Silander, J. A. 1998. A wavelet transform method to merge Landsat TM and SPOT panchromatic data. International Journal of Remote Sensing, 19: 743--757

1998

Pith tools

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

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Y.; Nam, M.-R.; and Kim, H

Choi, M.; Kim, R. Y.; Nam, M.-R.; and Kim, H. O. 2005. Fusion of multispectral and panchromatic Satellite images using the curvelet transform. IEEE Geoscience and Remote Sensing Letters, 2(2): 136--140

work page 2005

[4] [4]

W.; Graham, M.; Poorthuis, A.; Shelton, T.; Stephens, M.; Wilson, M

Crampton, J. W.; Graham, M.; Poorthuis, A.; Shelton, T.; Stephens, M.; Wilson, M. W.; and Zook, M. 2013. Beyond the geotag: Situating ‘big data’ and leveraging the potential of the geoweb. Cartography and Geographic Information Science, 40(2): 130--139

work page 2013

[5] [5]

L.; Denker, J

Cun, Y. L.; Denker, J. S.; and Solla, S. A. 1989. Optimal Brain Damage. In NIPS'89: Proceedings of the 3rd International Conference on Neural Information Processing Systems, 598--605. Association for Computing Machinery

work page 1989

[6] [6]

Deng, L.-J.; Vivone, G.; Jin, C.; and Chanussot, J. 2021. Detail Injection-Based Deep Convolutional Neural Networks for Pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 59(8): 6995--7010

work page 2021

[7] [7]

Fitzner, D.; Sester, M.; Haberlandt, U.; and Rabiei, E. 2013. Rainfall estimation with a geosensor network of cars: Theoretical considerations and first results. Photogrammetrie, Fernerkundung, Geoinformation, 2013(2): 93--103

work page 2013

[8] [8]

Garzelli, A.; and Nencini, F. 2009. Hypercomplex Quality Assessment of Multi/Hyperspectral Images. IEEE Geoscience and Remote Sensing Letters, 6(4): 662--665

work page 2009