pith. sign in

arxiv: 2606.23356 · v1 · pith:AVHETJSMnew · submitted 2026-06-22 · 💻 cs.CV · cs.LG

Changing Modalities: Adapting Remote Sensing Models to New Satellites and Sensors

Tim G. Zhou , Anthony Fuller , Geoff Pleiss , Evan Shelhamer This is my paper

Pith reviewed 2026-06-26 09:12 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords remote sensingmodality adaptationmodality hallucinationsatellite sensorsmodel adaptationmulti-modal learningsensor substitution
0
0 comments X

The pith

DeluluNet trains remote sensing models on one set of sensors so they can still predict when sensors are swapped, added, or removed.

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

Remote sensing models are trained on fixed sensor inputs, yet new satellites introduce different sensors and old ones are retired. The paper defines three practical change scenarios: replacing one modality with another, adding modalities, or using only a subset. DeluluNet solves these by training a multi-modal architecture end-to-end from a unimodal teacher plus unlabeled data, learning to generate representations for any missing sensors. The resulting model continues to produce predictions without new labels or full retraining. A reader would care because satellite fleets evolve and relabeling every change is costly.

Core claim

DeluluNet is an architecture with modular components for modality transfer, addition, and peeking; it is trained end-to-end by modality hallucination, in which missing-modality representations are predicted from the modalities that are present, using a unimodal teacher and unlabeled multimodal data, so that the model maintains prediction capability when input modalities change.

What carries the argument

Modality hallucination: the process of predicting representations for absent sensors from those that remain, learned from a unimodal teacher on unlabeled multimodal data.

If this is right

  • A model trained on one satellite's sensors can be deployed on a replacement sensor without new labeled data.
  • Additional sensors can be incorporated into an existing model by hallucinating their representations during inference.
  • A model can operate on only a subset of its original sensors by hallucinating the missing ones.
  • The same training procedure supplies a single set of weights usable across all three change scenarios.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach could lengthen the operational life of models trained on older satellite generations.
  • It might reduce repeated data collection and labeling campaigns each time a sensor fleet is upgraded.
  • The same hallucination training pattern could be tried on other multi-sensor domains such as medical imaging or autonomous driving.

Load-bearing premise

Representations produced by hallucination remain accurate enough that prediction performance does not drop sharply when modalities change.

What would settle it

A controlled test on held-out remote sensing data in which DeluluNet is evaluated on each of the three modality-change scenarios and accuracy falls by more than a small margin relative to a retrained baseline.

Figures

Figures reproduced from arXiv: 2606.23356 by Anthony Fuller, Evan Shelhamer, Geoff Pleiss, Tim G. Zhou.

Figure 2
Figure 2. Figure 2: DeluluNet Architecture and Training Strategy. A. Shallow modality-specific transformers extract low-level information from available modalities; B. Masking and fusion transformers facilitate cross-modality fusion through masked image modeling, and enables prediction with incomplete input modalities at test time; C. Predictor heads learning switches between distilling from uni-modal f0 teacher on Da and lea… view at source ↗
Figure 4
Figure 4. Figure 4: Panopticon Zero-Shot vs. Delulu for transfer. DeluluNet takes advantage of widely available unlabeled paired multi-modal data, achieving significantly better trans￾fer performance. modal model from a uni-modal teacher, aiming to make better predictions. We found that while MKE can sometimes outperform the uni-modal f0, DeluluNet provides more consistent and sizable improvements [PITH_FULL_IMAGE:figures/fu… view at source ↗
Figure 5
Figure 5. Figure 5: (left) Patch-level Pearson Correlation between real and hallucinated patch representa￾tions (quantitative). Despite low correlation between real S1 and S2 tokens, masking transformers achieve high linear correlation between real and hallucinated patch tokens across both modalities. (right) DeluluNet modality hallucinations (qualitative). Top-3 PCA components visualized as RGB show high similarity between r… view at source ↗
Figure 6
Figure 6. Figure 6: DeluluNet benefits from more allocation of layers for modality fusion: earlier fusion tends to lead to bet￾ter results across transfer, peeking, and addition. For modality transfer, earlier fusion even leads to transfer results that surpass DINOv3 supervised fine-tuned on Sentinel-1 data. data when the input modality is incomplete. This is achievable for DeluluNet because of masking transformers’ ability t… view at source ↗
Figure 7
Figure 7. Figure 7: Label-Free Validation Vs. Test metric (left) peeking uses standard validation evaluation since we have labels for MA. (center&right) Transfer and Addition uses their agreement with teacher f0 times peeking performance as val metric, providing a validation score that doesn’t require labels for multimodal validation split, and corresponds to test split metric reasonably well. There are many ways to define “a… view at source ↗
read the original abstract

Machine learning models for remote sensing are trained and deployed on a static set of modalities. However, as we equip newer satellites with novel sensors and retire old ones, practitioners may wish to deploy an existing model on a substitution, superset, or subset of modalities with minimal retraining given data availability or practical computational constraints. We study the setting of updating existing models to changing modalities and identify three main scenarios: Modality Transfer (substitution), Addition (superset), and Peeking (subset). We propose DeluluNet, an architecture with modular components for all three changing modality scenarios. DeluluNet is trained end-to-end, learning a multi-modal model from a unimodal teacher and unlabeled multimodal data via modality hallucination--predicting missing modality representations from those that are present. As a result, DeluluNet can keep predicting even when input modalities change, providing a practical alternative to re-labeling and re-training in a changing world.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript introduces DeluluNet, a modular architecture for adapting remote sensing models to changing input modalities without full retraining. It identifies three scenarios—Modality Transfer (substitution), Addition (superset), and Peeking (subset)—and trains the model end-to-end via modality hallucination, using a unimodal teacher and unlabeled multimodal data to predict missing modality representations.

Significance. If the central claims hold with supporting evidence, the work would address a practical deployment challenge in remote sensing where satellites and sensors evolve over time. By framing the problem around three concrete scenarios and proposing a hallucination-based alternative to re-labeling, it could reduce computational and data costs for model updates.

major comments (2)
  1. [Abstract] Abstract: The practicality claim—that DeluluNet provides an alternative to re-training when modalities change—depends on the availability of unlabeled multimodal data with co-occurring samples from old and new modalities to learn the hallucinator. For novel sensors on new satellites (where legacy satellites may be retired), such paired data often does not exist, and no zero-shot mechanism for unseen sensor distributions is described.
  2. [Abstract] Abstract: The assertion that modality hallucination 'can keep predicting even when input modalities change' without large performance drops is presented as a result, yet the provided text contains no experimental results, ablation studies, or quantitative metrics across the three scenarios to verify that the learned representations support effective downstream prediction.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for reviewing our manuscript and providing these insightful comments. We address each major comment below and outline the revisions we plan to make.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The practicality claim—that DeluluNet provides an alternative to re-training when modalities change—depends on the availability of unlabeled multimodal data with co-occurring samples from old and new modalities to learn the hallucinator. For novel sensors on new satellites (where legacy satellites may be retired), such paired data often does not exist, and no zero-shot mechanism for unseen sensor distributions is described.

    Authors: We concur that the proposed approach presupposes the existence of unlabeled multimodal data containing co-occurring samples from both the original and new modalities. This data is necessary to train the modality hallucinator. Our work focuses on the practical setting where such data can be acquired during periods of sensor transition. We do not propose or evaluate a zero-shot approach for entirely novel sensor distributions without any paired samples, which remains an open challenge. In the revised version, we will explicitly articulate this assumption in the abstract and add a discussion of this limitation. revision: partial

  2. Referee: [Abstract] Abstract: The assertion that modality hallucination 'can keep predicting even when input modalities change' without large performance drops is presented as a result, yet the provided text contains no experimental results, ablation studies, or quantitative metrics across the three scenarios to verify that the learned representations support effective downstream prediction.

    Authors: The manuscript body contains the experimental section with results, ablations, and metrics for the three scenarios. The abstract condenses these outcomes. To improve clarity, we will revise the abstract to incorporate key quantitative findings, such as the observed performance metrics under modality changes. revision: yes

Circularity Check

0 steps flagged

No circularity detected in DeluluNet derivation

full rationale

The paper introduces DeluluNet as a new modular architecture trained end-to-end via modality hallucination from a unimodal teacher and unlabeled multimodal data. This supports the three scenarios (transfer, addition, peeking) through the described training procedure without any load-bearing steps that reduce by construction to fitted parameters, self-definitions, or self-citation chains. The central claim of providing a practical alternative to re-labeling remains independent of its inputs and is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Based solely on the abstract, the paper introduces the DeluluNet architecture and the modality-hallucination training objective but provides no explicit free parameters, axioms, or invented physical entities.

invented entities (1)
  • DeluluNet no independent evidence
    purpose: Modular architecture supporting modality transfer, addition, and peeking via hallucination
    New model name and design introduced in the abstract

pith-pipeline@v0.9.1-grok · 5699 in / 1103 out tokens · 26670 ms · 2026-06-26T09:12:29.598882+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 7 canonical work pages

  1. [1]

    Omnisat: Self- supervised modality fusion for earth observation

    Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Omnisat: Self- supervised modality fusion for earth observation. InEuropean Conference on Computer Vision, pages 409–427. Springer, 2024

    2024

  2. [2]

    Mixmatch: A holistic approach to semi-supervised learning.Advances in neural information processing systems, 32, 2019

    David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning.Advances in neural information processing systems, 32, 2019

    2019

  3. [3]

    Combining labeled and unlabeled data with co-training

    Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the Eleventh Annual Conference on Computational Learning Theory, COLT’ 98, page 92–100, New York, NY , USA, 1998. Association for Computing Machinery. ISBN 1581130570. doi: 10.1145/279943.279962

    work page doi:10.1145/279943.279962 1998

  4. [4]

    Model compression , isbn =

    Cristian Bucila, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. volume 2006, pages 535–541, 08 2006. doi: 10.1145/1150402.1150464

    work page doi:10.1145/1150402.1150464 2006

  5. [5]

    Bootstrapping chest ct image understanding by distilling knowledge from x-ray expert models

    Weiwei Cao, Jianpeng Zhang, Yingda Xia, Tony CW Mok, Zi Li, Xianghua Ye, Le Lu, Jian Zheng, Yuxing Tang, and Ling Zhang. Bootstrapping chest ct image understanding by distilling knowledge from x-ray expert models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11238–11247, 2024

    2024

  6. [6]

    Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning

    Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. InProceedings of the AAAI conference on artificial intelligence, volume 35, pages 6912–6920, 2021

    2021

  7. [7]

    reben: Refined bigearthnet dataset for remote sensing image analysis

    Kai Norman Clasen, Leonard Hackel, Tom Burgert, Gencer Sumbul, Begüm Demir, and V olker Markl. reben: Refined bigearthnet dataset for remote sensing image analysis. InIGARSS 2025-2025 IEEE International Geoscience and Remote Sensing Symposium, pages 1264–1268. IEEE, 2025

    2025

  8. [8]

    Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery.Advances in Neural Information Processing Systems, 35: 197–211, 2022

    Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery.Advances in Neural Information Processing Systems, 35: 197–211, 2022

    2022

  9. [9]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representations, 2021

    2021

  10. [10]

    and Del Bello, U

    M. Drusch, U. Del Bello, S. Carlier, O. Colin, V . Fernandez, F. Gascon, B. Hoersch, C. Isola, P. Laberinti, P. Martimort, A. Meygret, F. Spoto, O. Sy, F. Marchese, and P. Bargellini. Sentinel- 2: Esa’s optical high-resolution mission for gmes operational services.Remote Sensing of Environment, 120:25–36, 2012. ISSN 0034-4257. doi: https://doi.org/10.1016...

    work page doi:10.1016/j.rse.2011.11.026 2012

  11. [11]

    URL https://www.esa.int/Applications/Observing_the_Earth/ Copernicus/Free_access_to_Copernicus_Sentinel_satellite_data

    ESA, 2013. URL https://www.esa.int/Applications/Observing_the_Earth/ Copernicus/Free_access_to_Copernicus_Sentinel_satellite_data. Accessed: 2026-02-04. 10

    2013

  12. [12]

    Tessera: Temporal embeddings of surface spectra for earth representation and analysis, 2025

    Zhengpeng Feng et al. Tessera: Temporal embeddings of surface spectra for earth representation and analysis, 2025. URLhttps://arxiv.org/abs/2506.20380

    Pith/arXiv arXiv 2025

  13. [13]

    Clay foundation model, 2025

    Clay Foundation. Clay foundation model, 2025. URL https://clay-foundation.github. io/model/index.html. Accessed: 2025-07-21

    2025

  14. [14]

    Croma: Remote sensing representations with contrastive radar-optical masked autoencoders.Advances in Neural Information Processing Systems, 36:5506–5538, 2023

    Anthony Fuller, Koreen Millard, and James Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders.Advances in Neural Information Processing Systems, 36:5506–5538, 2023

    2023

  15. [15]

    Domain-adversarial training of neural networks

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1–35, 2016

    2016

  16. [16]

    Goward, Laura E

    Samuel N. Goward, Laura E. P. Rocchio, Darrel L. Williams, Terry Arvidson, James R. Irons, Carol A. Russell, and Shaida S. Johnston. Landsat’s enduring legacy: Pioneering global land observations from space.Photogrammetric Engineering and Remote Sensing, 84:9–10, 2017. URLhttps://api.semanticscholar.org/CorpusID:134711498

    2017

  17. [17]

    Semi-supervised learning by entropy minimization

    Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In L. Saul, Y . Weiss, and L. Bottou, editors,Advances in Neural Information Processing Systems, volume 17. MIT Press, 2004. URL https://proceedings.neurips.cc/paper_files/ paper/2004/file/96f2b50b5d3613adf9c27049b2a888c7-Paper.pdf

    2004

  18. [18]

    Cross modal distillation for supervision transfer

    Saurabh Gupta, Judy Hoffman, and Jitendra Malik. Cross modal distillation for supervision transfer. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2827–2836, 2016

    2016

  19. [19]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000–16009, 2022

    2022

  20. [20]

    Modality-incremental learning with disjoint relevance mapping networks for image-based semantic segmentation

    Niharika Hegde, Shishir Muralidhara, René Schuster, and Didier Stricker. Modality-incremental learning with disjoint relevance mapping networks for image-based semantic segmentation. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5540–5549. IEEE, 2025

    2025

  21. [21]

    Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217–2226, 2019

    2019

  22. [22]

    A comprehensive overhaul of feature distillation

    Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. InProceedings of the IEEE/CVF international conference on computer vision, pages 1921–1930, 2019

    1921

  23. [23]

    Olmoearth: Stable latent image modeling for multimodal earth observation, 2025

    Henry Herzog, Favyen Bastani, Yawen Zhang, Gabriel Tseng, Joseph Redmon, Hadrien Sablon, Ryan Park, Jacob Morrison, Alexandra Buraczynski, Karen Farley, Joshua Hansen, Andrew Howe, Patrick Alan Johnson, Mark Otterlee, Ted Schmitt, Hunter Pitelka, Stephen Daspit, Rachel Ratner, Christopher Wilhelm, Sebastian Wood, Mike Jacobi, Hannah Kerner, Evan Shelhamer...

    2025

  24. [24]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    Pith/arXiv arXiv 2015

  25. [25]

    Mae- stro: Masked autoencoders for multimodal, multitemporal, and multispectral earth observation data, 2025

    Antoine Labatie, Michael Vaccaro, Nina Lardiere, Anatol Garioud, and Nicolas Gonthier. Mae- stro: Masked autoencoders for multimodal, multitemporal, and multispectral earth observation data, 2025. URLhttps://arxiv.org/abs/2508.10894

    arXiv 2025

  26. [26]

    Geo-bench: Toward foundation models for earth monitoring.Advances in Neural Information Processing Systems, 36:51080–51093, 2023

    Alexandre Lacoste, Nils Lehmann, Pau Rodriguez, Evan Sherwin, Hannah Kerner, Björn Lütjens, Jeremy Irvin, David Dao, Hamed Alemohammad, Alexandre Drouin, et al. Geo-bench: Toward foundation models for earth monitoring.Advances in Neural Information Processing Systems, 36:51080–51093, 2023. 11

    2023

  27. [27]

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks

    Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. InWorkshop on challenges in representation learning, ICML, 2013

    2013

  28. [28]

    Why m heads are better than one: Training a diverse ensemble of deep networks.arXiv preprint arXiv:1511.06314, 2015

    Stefan Lee, Senthil Purushwalkam, Michael Cogswell, David Crandall, and Dhruv Batra. Why m heads are better than one: Training a diverse ensemble of deep networks.arXiv preprint arXiv:1511.06314, 2015

    Pith/arXiv arXiv 2015

  29. [29]

    Rethinking the stability-plasticity trade- off in continual learning from an architectural perspective

    Aojun Lu, Hangjie Yuan, Tao Feng, and Yanan Sun. Rethinking the stability-plasticity trade- off in continual learning from an architectural perspective. InForty-second International Conference on Machine Learning, 2025

    2025

  30. [30]

    Schutz, Benjamin Smith, Yuekui Yang, and Jay Zwally

    Thorsten Markus, Tom Neumann, Anthony Martino, Waleed Abdalati, Kelly Brunt, Beata Csatho, Sinead Farrell, Helen Fricker, Alex Gardner, David Harding, Michael Jasinski, Ron Kwok, Lori Magruder, Dan Lubin, Scott Luthcke, James Morison, Ross Nelson, Amy Neuenschwander, Stephen Palm, Sorin Popescu, CK Shum, Bob E. Schutz, Benjamin Smith, Yuekui Yang, and Jay...

    work page doi:10.1016/j.rse.2016.12.029 2017

  31. [31]

    Michael McCloskey and Neal J. Cohen. Catastrophic interference in connectionist net- works: The sequential learning problem. In Gordon H. Bower, editor,Psychology of Learn- ing and Motivation, volume 24 ofPsychology of Learning and Motivation, pages 109–

  32. [32]

    1989 , issn =

    Academic Press, 1989. doi: https://doi.org/10.1016/S0079-7421(08)60536-8. URL https://www.sciencedirect.com/science/article/pii/S0079742108605368

    work page doi:10.1016/s0079-7421(08)60536-8 1989

  33. [33]

    4m: Massively multimodal masked modeling.Advances in Neural Information Processing Systems, 36:58363–58408, 2023

    David Mizrahi, Roman Bachmann, Oguzhan Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4m: Massively multimodal masked modeling.Advances in Neural Information Processing Systems, 36:58363–58408, 2023

    2023

  34. [34]

    Hyperspectral imaging technologies, 2019

    NRC. Hyperspectral imaging technologies, 2019. URL https://nrc.canada.ca/ en/research-development/products-services/technical-advisory-services/ hyperspectral-imaging-technologies. Accessed: 2026-02-04

    2019

  35. [35]

    xview3-sar: Detecting dark fishing activity using synthetic aperture radar imagery.Advances in Neural Information Processing Systems, 35: 37604–37616, 2022

    Fernando Paolo, Tsu-ting Tim Lin, Ritwik Gupta, Bryce Goodman, Nirav Patel, Daniel Kuster, David Kroodsma, and Jared Dunnmon. xview3-sar: Detecting dark fishing activity using synthetic aperture radar imagery.Advances in Neural Information Processing Systems, 35: 37604–37616, 2022

    2022

  36. [36]

    Teacher-student domain adaptation for biosensor models.ICLR, 2020

    Lawrence G Phillips, David B Grimes, and Yihan Jessie Li. Teacher-student domain adaptation for biosensor models.ICLR, 2020

    2020

  37. [37]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PMLR, 2021

    2021

  38. [38]

    Monitrs: Multimodal observations of natural incidents through remote sensing.arXiv preprint arXiv:2507.16228, 2025

    Shreelekha Revankar, Utkarsh Mall, Cheng Perng Phoo, Kavita Bala, and Bharath Hariharan. Monitrs: Multimodal observations of natural incidents through remote sensing.arXiv preprint arXiv:2507.16228, 2025

    arXiv 2025

  39. [39]

    Caleb Robinson, Kolya Malkin, Nebojsa Jojic, Huijun Chen, Rongjun Qin, Changlin Xiao, Michael Schmitt, Pedram Ghamisi, Ronny Hänsch, and Naoto Yokoya. Global land-cover mapping with weak supervision: Outcome of the 2020 ieee grss data fusion contest.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:3185–3199,

    2020

  40. [40]

    doi: 10.1109/JSTARS.2021.3063849

    work page doi:10.1109/jstars.2021.3063849 2021

  41. [41]

    Oriane Siméoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julie...

    Pith/arXiv arXiv 2025

  42. [42]

    Geo- bench-2: From performance to capability, rethinking evaluation in geospatial ai.arXiv preprint arXiv:2511.15658, 2025

    Naomi Simumba, Nils Lehmann, Paolo Fraccaro, Hamed Alemohammad, Geeth De Mel, Salman Khan, Manil Maskey, Nicolas Longepe, Xiao Xiang Zhu, Hannah Kerner, et al. Geo- bench-2: From performance to capability, rethinking evaluation in geospatial ai.arXiv preprint arXiv:2511.15658, 2025

    arXiv 2025

  43. [43]

    Fixmatch: Simplifying semi- supervised learning with consistency and confidence.Advances in neural information processing systems, 33:596–608, 2020

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raf- fel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi- supervised learning with consistency and confidence.Advances in neural information processing systems, 33:596–608, 2020

    2020

  44. [44]

    Bigearthnet: A large-scale benchmark archive for remote sensing image understanding.arXiv preprint arXiv:1902.06148, 2019

    Gencer Sumbul, Marcela Charfuelan, Begüm Demir, and V olker Markl. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding.arXiv preprint arXiv:1902.06148, 2019

    arXiv 1902

  45. [45]

    Deep coral: Correlation alignment for deep domain adaptation

    Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. InEuropean conference on computer vision, pages 443–450. Springer, 2016

    2016

  46. [46]

    Geocrossbench: Cross-band generalization for remote sensing.arXiv preprint arXiv:2511.02831, 2025

    Hakob Tamazyan, Ani Vanyan, Alvard Barseghyan, Anna Khosrovyan, Evan Shelhamer, and Hrant Khachatrian. Geocrossbench: Cross-band generalization for remote sensing.arXiv preprint arXiv:2511.02831, 2025

    arXiv 2025

  47. [47]

    Galileo: Learning global & local features of many remote sensing modalities

    Gabriel Tseng, Anthony Fuller, Marlena Reil, Henry Herzog, Patrick Beukema, Favyen Bastani, James R Green, Evan Shelhamer, Hannah Kerner, and David Rolnick. Galileo: Learning global & local features of many remote sensing modalities. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Z...

    2025

  48. [48]

    Galileo: Learning global & local features of many remote sensing modalities

    Gabriel Tseng, Anthony Fuller, Marlena Reil, Henry Herzog, Patrick Beukema, Favyen Bastani, James R Green, Evan Shelhamer, Hannah Kerner, and David Rolnick. Galileo: Learning global & local features of many remote sensing modalities. InInternational Conference on Machine Learning, pages 60280–60300. PMLR, 2025

    2025

  49. [49]

    Three types of incremental learning.Nature Machine Intelligence, 4(12):1185–1197, 2022

    Gido M Van de Ven, Tinne Tuytelaars, and Andreas S Tolias. Three types of incremental learning.Nature Machine Intelligence, 4(12):1185–1197, 2022

    2022

  50. [50]

    Panopticon: Advancing any-sensor foundation models for earth observation

    Leonard Waldmann, Ando Shah, Yi Wang, Nils Lehmann, Adam Stewart, Zhitong Xiong, Xiao Xiang Zhu, Stefan Bauer, and John Chuang. Panopticon: Advancing any-sensor foundation models for earth observation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 2204–2214, 2025

    2025

  51. [51]

    Freematch: Self-adaptive thresholding for semi-supervised learning

    Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. InEleventh International Conference on Learning Representations. OpenReview. net, 2023

    2023

  52. [52]

    Neural plasticity- inspired multimodal foundation model for earth observation.arXiv preprint arXiv:2403.15356, 2024

    Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity- inspired multimodal foundation model for earth observation.arXiv preprint arXiv:2403.15356, 2024

    arXiv 2024

  53. [53]

    One for all: Toward unified foundation models for earth vision

    Zhitong Xiong, Yi Wang, Fahong Zhang, and Xiao Xiang Zhu. One for all: Toward unified foundation models for earth vision. InIGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, pages 2734–2738, 2024. doi: 10.1109/IGARSS53475.2024. 10641637

    work page doi:10.1109/igarss53475.2024 2024

  54. [54]

    Multimodal knowledge expansion

    Zihui Xue, Sucheng Ren, Zhengqi Gao, and Hang Zhao. Multimodal knowledge expansion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 854–863, 2021

    2021

  55. [55]

    mixup: Beyond empirical risk minimization

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. InInternational Conference on Learning Representations, 2018. 13

    2018

  56. [56]

    Skysense v2: A unified foundation model for multi-modal remote sensing

    Yingying Zhang, Lixiang Ru, Kang Wu, Lei Yu, Lei Liang, Yansheng Li, and Jingdong Chen. Skysense v2: A unified foundation model for multi-modal remote sensing. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 9136–9146, 2025

    2025

  57. [57]

    Through-wall human pose estimation using radio signals

    Mingmin Zhao, Tianhong Li, Mohammad Abu Alsheikh, Yonglong Tian, Hang Zhao, Antonio Torralba, and Dina Katabi. Through-wall human pose estimation using radio signals. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    2018

  58. [58]

    Knowledge distillation based on transformed teacher matching.arXiv preprint arXiv:2402.11148, 2024

    Kaixiang Zheng and En-Hui Yang. Knowledge distillation based on transformed teacher matching.arXiv preprint arXiv:2402.11148, 2024

    arXiv 2024

  59. [59]

    Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird’s-eye view

    Shengchao Zhou, Weizhou Liu, Chen Hu, Shuchang Zhou, and Chao Ma. Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird’s-eye view. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5116–5125, 2023

    2023

  60. [60]

    agreement

    Tim G Zhou, Evan Shelhamer, and Geoff Pleiss. Asymmetric duos: Sidekicks improve uncertainty. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 14 A Label-Free Early Stopping for DeluluNet Transfer and Addition Similar to training splits Dl and Da, we assume a labeled uni-modal MA,Y validation set and an unlabeled multi-...

    2025