Pith. sign in

REVIEW 3 major objections 6 minor 31 references

SPECTRA shows that geospatial foundation models can use every available sensor band, not just wavelength-matched ones, while fine-tuning with only 3.3–7.7% of parameters.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A fine-tuning framework that uses all available satellite bands via a residual gated adapter and allocates LoRA ranks by stage-level transferability, improving cross-sensor segmentation at lower parameter cost.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Useful BRE result with real experimental breadth, but the ST-LoRA component is not actually isolated from BRE, and the evidence for it is thin enough that the paper's second half needs either new controls or much softer claims. the 3 major comments →

arxiv 2608.01751 v1 pith:CCB5TG34 submitted 2026-08-03 cs.CV cs.AI

SPECTRA: Band-Routed Embedding and Stage-Wise LoRA for Cross-Sensor Fine-Tuning of Geospatial Foundation Models

Xingyan Li , Jordan A. Caraballo-Vega , Jie Gong , Mark L. Carroll , Jianwu Wang This is my paper
classification cs.CV cs.AI
keywords geospatial foundation modelscross-sensor fine-tuningspectral band mismatchBand-Routed Embeddingstage-wise LoRALogME transferabilityparameter-efficient fine-tuningsemantic segmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 argues that two standard problems in adapting Earth-observation foundation models—spectral band mismatch and high fine-tuning cost—can be solved together. Its framework, SPECTRA, combines Band-Routed Embedding (BRE) with Stage-wise Transferability-aware LoRA (ST-LoRA). BRE starts from the conventional band-selection input and adds a zero-initialized, gated residual correction that lets every downstream channel influence every pretrained input band without modifying the pretrained patch embedding. ST-LoRA measures each encoder stage's transferability with LogME before training and redistributes a fixed LoRA rank budget toward stages that should help the task. Across three pretrained models, four segmentation datasets, and five fine-tuning policies, adding BRE improves mean macro mIoU by +2.88 points over matched band-selection baselines, with ST-LoRA keeping trainable parameters at 3.29–7.73%.

Core claim

SPECTRA's central claim is that spectral adaptation and parameter-efficient fine-tuning are compatible. The BRE input adapter maps all C observed channels into the K-band space the pretrained patch embedding expects; because the residual path is zero-initialized, the first forward pass exactly equals band-selection, and training learns routing-gated contributions from all bands without changing the pretrained tokenizer. The paper reports that this design improves mean macro mIoU by 2.88 points over the matched band-selection input across 60 backbone–dataset–policy configurations, while a direct all-band MLP projector lowers it by 2.28 points. ST-LoRA then controls cost: it computes a per-sta

What carries the argument

Band-Routed Embedding (BRE) is a zero-initialized residual adapter with a K×C sigmoid gate table: each downstream band i is gated into each virtual pretrained band j, all K×C gated channels are concatenated, passed through one shared 1×1-convolution residual adapter, and added to the band-selection anchor Xsel. The anchor keeps the pretrained patch embedding valid, the zero initialization makes BRE initialize as band-selection, and the gating guarantees no target channel is hard-discarded. Stage-wise Transferability-aware LoRA (ST-LoRA) is a one-time rank planner: it runs LogME on frozen features per encoder stage, normalizes transfer or repair scores with a temperature softmax, quantizes co

Load-bearing premise

The planner assumes that a LogME score computed once on frozen, band-selected features predicts which encoder stages will gain most from LoRA adaptation, and that assumption is validated on only one of the four datasets.

What would settle it

On a dataset other than Sen1Floods11, compute the ST-LoRA rank schedule, then train its exact reverse (the stage given the most rank gets 0, the stage given 0 gets the most); if the reverse schedule matches or beats the planned schedule, the LogME transferability ranking is not what carries ST-LoRA's gains.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Any fixed-interface pretrained GeoFM could accept arbitrary multisensor inputs without retraining its patch embedding, making older backbones reusable for new sensors.
  • The average +2.88 mIoU and +3.08 F1 gains indicate that band-selection discards task-relevant spectral information, so all-band input adaptation is preferable whenever the target sensor has more channels than the pretrained interface.
  • Transferability diagnostics computed once on frozen features can be used to commit LoRA rank budgets before training, avoiding sweeps over uniform ranks.
  • A user can replace uniform LoRA-32 with ST-LoRA under the same total rank budget and use fewer trainable parameters with similar or better accuracy.
  • The repair planning mode suggests that, in some settings, stages with low transferability need more corrective capacity, not less.

Where Pith is reading between the lines

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

  • The BRE design principle—anchor to the pretrained input distribution and learn a zero-initialized residual from all channels—could generalize beyond geospatial data to any fixed-embedding model receiving heterogeneous input modalities, though the paper does not claim this.
  • A cheaper alternative to LogME would be a small linear probe per stage on the same frozen features; if probe rankings agree with LogME, the planner's success is about linear usefulness, and if they disagree, the choice of transferability metric becomes decisive.
  • The tables show SPECTRA sometimes exceeding full fine-tuning; if that pattern reproduces outside the 12 evaluated combinations, it would suggest full fine-tuning is wasteful when the pretrained representation is already well aligned, but the paper only claims a favorable accuracy-cost tradeoff.
  • Because BRE concatenates K×C gated channels before its 1×1 convolution, peak memory grows with the number of observed bands; reporting peak GPU memory alongside trainable parameters would clarify the practical cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes SPECTRA, a parameter-efficient fine-tuning framework for cross-sensor adaptation of geospatial foundation models. It addresses spectral mismatch via Band-Routed Embedding (BRE), which starts from a selected-band anchor and adds a zero-initialized, gated residual path that maps all available downstream bands to the pretrained-compatible band space. It addresses adaptation cost via Stage-wise Transfer-aware LoRA (ST-LoRA), which uses LogME computed on frozen selected-band features to allocate LoRA ranks across encoder stages. Experiments span three GeoFMs (Prithvi-EO-2.0, ScaleMAE, SatMAE) and four segmentation datasets. The core quantitative claim is that BRE improves mean macro mIoU by +2.88 points over band-selection across 60 matched backbone–dataset–policy cells (Table 5), while SPECTRA uses only 3.29–7.73% trainable parameters. The ST-LoRA evidence is confined to Sen1Floods11 and lacks a matched uniform-LoRA/BRE control.

Significance. The BRE contribution is well-motivated and supported by a solid matched ablation: Table 5 shows consistent positive gains over band-selection across many configurations, and the zero-initialized anchor design is elegant and practical. The paper provides code, seed-level statistics, and a clear conditional interpretation of when extra-band adapters help. The ST-LoRA contribution, however, is not yet established. The current evidence is limited to one dataset, the rank-planning proxy assumption is untested, and the missing control row prevents attribution of any gain to the rank planner. If ST-LoRA were dropped or reframed, the BRE result alone would be a useful incremental contribution to cross-sensor GeoFM fine-tuning.

major comments (3)
  1. [Section 5.4, Table 6] The ST-LoRA component is not isolated. The text references a 'uniform LoRA-32 BRE reference' but this row is not reported in Table 6. Manual schedules M1–M8 do not include a uniform allocation under BRE, and their spread overlaps with ST-LoRA (e.g., Prithvi M1 86.58±0.71 vs ST-LoRA 85.88±1.24; ScaleMAE M1 85.27±1.77 vs ST-LoRA 85.29±2.17). Moreover, Table 6 does not state the input adapter used for M1–M8 or the transfer/repair mode of the ST-LoRA row; the ST-LoRA row for Prithvi (85.88) does not match the SPECTRA result in Table 4 (86.39), so the row is not the configuration reported in the main table. Without the matched control, the second core contribution is unvalidated.
  2. [Section 4.2, Eq. (7)-(8)] The ST-planner computes LogME on features from the selected-band path (Eq. 7), but fine-tuning uses BRE-modified inputs. The proxy assumption that rank transferability measured under band-selection predicts LoRA benefit under BRE is not stated or tested. Validation is limited to Sen1Floods11, as Section 6 concedes: 'broader rank-planner validation across all target datasets remains necessary.' If LogME does not rank stages by LoRA benefit, ST-LoRA is a heuristic with no demonstrated benefit. At minimum, a matched BRE+uniform-LoRA control and one additional dataset would be needed.
  3. [Abstract; Table 4] The abstract claims SPECTRA 'reduces trainable parameters compared with full fine-tuning and standard LoRA' without specifying the reference rank. In Table 4, SPECTRA's trainable percentage exceeds LoRA-16 in several cells (e.g., ScaleMAE FireScars 7.48% vs 6.73%; SatMAE FireScars 7.73% vs 6.73%). The efficiency claim should either specify the LoRA rank used for comparison (presumably LoRA-32, as in Section 5.4) or be softened. This does not affect the BRE evidence, but it affects the stated contribution.
minor comments (6)
  1. [Section 4.1] Several typos and placeholder artifacts need proofreading: 'inpu band', 'fo the virtual image', 'gated paris', 'xxx embe- ded tokens', 'learns xx'.
  2. [Section 4.2] The temperature τ and the patch-purity threshold are free hyperparameters; please report their values and provide sensitivity analysis, since the rank allocation depends on them.
  3. [Table 6] Please add a 'uniform LoRA-32 BRE' row, label the ST-LoRA rows with their mode (transfer/repair), and state the input adapter used for each manual schedule.
  4. [Table 3] ScaleMAE Landslide4Sense transfer gap is listed as 0.000; if this is a rounding artifact, clarify or report with more precision.
  5. [Table 4] Several cells have large standard deviations (e.g., ScaleMAE Sen1Floods11 LP 70.16±4.12); consider discussing this variability or reporting more seeds.
  6. [Section 5.1, Table 5] The 'direct MLP projector' baseline is not described in detail (architecture, number of parameters). Specify it so the comparison in Table 5 is fair and reproducible.

Circularity Check

0 steps flagged

No significant circularity: SPECTRA is an empirical method paper whose claims are evaluated on held-out test sets; the stated limitations are honest and do not hide a fitted-parameter-as-prediction or self-citation chain.

full rationale

This paper is an empirical method paper, not a derivation from first principles, and no central claim reduces by construction to its own inputs. The headline BRE result (+2.88 mean macro mIoU over 60 matched cells, Table 5) is a held-out test-set comparison between two input adapters while keeping the fine-tuning policy fixed; BRE starts from the band-selection anchor with a zero-initialized residual (Eq. 4), so the improvement is measured, not guaranteed by the definition of 'fully using bands.' The ST-LoRA rank planner (Section 4.2) uses LogME on frozen selected-band features (Eq. 8) to choose a stage-wise rank schedule before fine-tuning; the schedule is then evaluated on a held-out test set against manual schedules (Table 6). This is a proxy-based heuristic selection, not a fitted constant recycled as a prediction. Selecting the better of transfer/repair modes by validation performance is standard model selection on a validation split, not test-set fitting. There are no load-bearing self-citations: all cited prior work (LogME, LoRA, GEO-Bench, GeoFM backbones) is external to the present authors, and no uniqueness theorem from the authors' own prior work is invoked. The paper explicitly concedes in Section 6, 'broader rank-planner validation across all target datasets remains necessary before claiming universal parameter-efficiency gains,' and similar limitations are stated for BRE gains. These limitations are honest scope statements, not circularity. The skeptic's concern about the missing BRE+uniform-LoRA control is an experimental-isolation gap and a correctness risk, not circular reasoning: no equation or fitted value is being disguised as an independent result. Therefore the paper is self-contained against external benchmarks and earns a score of 0.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard assumptions about frozen pretrained interfaces, the quality of wavelength-based band selection, and the proxy validity of LogME for LoRA transferability. The main unquantified design choices are the temperature, the purity threshold, and the validation-guided planner mode selection. No new physical entities are introduced.

free parameters (3)
  • ST-LoRA softmax temperature tau
    Controls how concentrated the stage-wise rank allocation is (Eq. 16). The paper never reports the value used; it directly changes the resulting rank schedule and thus the accuracy/cost tradeoff.
  • LogME patch purity threshold
    Used in Section 4.2 to keep only patches whose dominant-class purity exceeds a threshold. The threshold value is not given; it determines the feature matrix and hence the transferability scores.
  • ST-LoRA mode (transfer vs repair) = selected by validation
    Reported SPECTRA results use the better of the two planning modes chosen by validation performance (Section 5.1). This categorical selection affects all reported SPECTRA numbers without being accounted for in the reported variances.
axioms (5)
  • domain assumption The pretrained GeoFM patch embedding expects a fixed K-band interface and is kept frozen during adaptation.
    Section 3.1 defines the problem around a frozen E0; BRE is designed to fit that interface. If one allowed retraining the embedding (as in any-sensor models), BRE would not be necessary.
  • domain assumption Band selection by central-wavelength matching is a fair, physically motivated baseline and the correct anchor for the adapter.
    Section 3.2 and Table 3 use wavelength matching to define S; BRE's selected-band anchor inherits this. The validity of BRE as a residual on top of this anchor depends on band-selection being a reasonable starting point.
  • domain assumption LogME computed on frozen selected-band features is a valid proxy for stage-wise transferability under LoRA fine-tuning.
    Section 4.2 uses LogME to plan rank allocation before fine-tuning. The paper validates this only on Sen1Floods11, and its own limitations section concedes broader validation is needed. If LogME does not rank stages by LoRA benefit, ST-LoRA collapses to a heuristic.
  • ad hoc to paper The discrete rank set R(r_ref) = {0,4,8,...,2r_ref} is an adequate design space for stage-wise ranks.
    Eq. (12) fixes this grid; it constrains the optimizer and the budget. The choice of grid granularity and the upper bound 2r_ref is a design decision made by the authors, not derived from first principles.
  • domain assumption A 1x1 convolutional residual adapter over concatenated gated band-pairs can capture task-relevant cross-band corrections without spatial context.
    Section 4.1 defines the residual adapter as a compact 1x1 conv net. The adapter has no spatial receptive field beyond the pixel, so any benefit from extra bands must be expressible as per-pixel spectral mixing. This is plausible but not proven; the paper does not ablate spatial kernels.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SPECTRA: Band-Routed Embedding and Stage-Wise LoRA for Cross-Sensor Fine-Tuning of Geospatial Foundation Models." pith.science (2026). https://pith.science/paper/CCB5TG34

@misc{pith2026260801751,
  author       = {Pith},
  title        = {Pith review of: SPECTRA: Band-Routed Embedding and Stage-Wise LoRA for Cross-Sensor Fine-Tuning of Geospatial Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CCB5TG34}},
  note         = {Machine review of arXiv:2608.01751}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Geospatial foundation models (GeoFMs), pretrained on large-scale geospatial data such as Earth observation (EO), climate, and weather data, have shown promising performance when fine-tuned on diverse downstream tasks. However, there are two challenges of adapting EO-pretrained GeoFMs to practical downstream datasets. The first challenge is how to handle spectral mismatch: pretrained patch embeddings expect a fixed set of input bands, whereas downstream sensors may provide different channels. The second challenge is how to reduce fine-tuning cost and make it efficient. While existing work has made efforts on these challenges individually, jointly improving fine-tuning performance under spectral mismatch while reducing adaptation cost remains underexplored. We propose SPECTRA, a parameter-efficient fine-tuning framework that addresses both spectral mismatch and adaptation cost. To handle spectral mismatch, SPECTRA introduces Band-Routed Embedding (BRE), which maps all available downstream bands into the band space expected by the pretrained GeoFM. By using BRE, all available bands in the downstream dataset are utilized to improve the selected-band input without changing the pretrained patch embedding interface. To reduce adaptation cost, SPECTRA further introduces a Stage-wise Transferability-aware LoRA (ST-LoRA) fine-tuning. ST-LoRA estimates stage-wise transferability before fine-tuning and assigns stage-specific LoRA ranks, concentrating trainable parameters on the stages with high transferability for the target task. Across three EO-pretrained GeoFMs and four downstream segmentation datasets, experiments show that BRE improves performance by utilizing all spectral bands, while ST-LoRA reduces trainable parameters compared with full fine-tuning and standard LoRA. Code is available at https://github.com/big-data-lab-umbc/SPECTRA.

Figures

Figures reproduced from arXiv: 2608.01751 by Jianwu Wang, Jie Gong, Jordan A. Caraballo-Vega, Mark L. Carroll, Xingyan Li.

Figure 1
Figure 1. Figure 1: Comparison of PEFT methods compared in this [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SPECTRA architecture. BRE maps a target image with [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of predictions for LoRA-32 band [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: BRE as a general input adapter across fine-tuning [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy-cost tradeoff curves across GeoFMs and [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

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

31 extracted references · 21 canonical work pages

  1. [1]

    Guillaume Astruc, Nicolas Gonthier, Clément Mallet, and Loïc Landrieu. 2025. AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modali- ties. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 19530–19540. doi:10.1109/CVPR52734.2025.01819

    arXiv 2025

  2. [2]

    Derrick Bonafilia, Beth Tellman, Tyler Anderson, and Erica Issenberg. 2020. Sen1Floods11: A Georeferenced Dataset to Train and Test Deep Learning Flood Algorithms for Sentinel-1. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 835–845

    work page 2020

  3. [3]

    Lobell, and Stefano Ermon

    Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David B. Lobell, and Stefano Ermon. 2022. SatMAE: Pre-training Transformers for Temporal and Multi-Spectral Satellite Imagery. InAdvances in Neural Information Processing Systems, Vol. 35. 197–211

    work page 2022

  4. [4]

    Zhe Dong, Yanfeng Gu, and Tianzhu Liu. 2024. UPetu: A Unified Parameter- Efficient Fine-Tuning Framework for Remote Sensing Foundation Model.IEEE Transactions on Geoscience and Remote Sensing62 (2024), 1–13. doi:10.1109/TGRS. 2024.3382734

    work page arXiv 2024

  5. [5]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    Pith/arXiv arXiv 2020

  6. [6]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. InInternational Conference on Learning Representations

    2021

  7. [7]

    Omid Ghorbanzadeh, Yonghao Xu, Pedram Ghamisi, Michael Kopp, and David Kreil. 2022. Landslide4Sense: Reference Benchmark Data and Deep Learning Models for Landslide Detection.IEEE Transactions on Geoscience and Remote Sensing60 (2022), 1–17. doi:10.1109/TGRS.2022.3215209

    work page arXiv 2022

  8. [8]

    2016.Deep Learning

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville. 2016.Deep Learning. MIT Press

    2016

  9. [9]

    Boran Han, Shuai Zhang, Xingjian Shi, and Markus Reichstein. 2024. Bridging Remote Sensors with Multisensor Geospatial Foundation Models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 27852– 27862. doi:10.1109/CVPR52733.2024.02631

    arXiv 2024

  10. [10]

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick

  11. [11]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-Efficient Transfer Learning for NLP. InProceedings of the 36th Interna- tional Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 97). PMLR, 2790–2799

    work page 2019

  12. [12]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations

    work page 2022

  13. [13]

    Phillips, Paolo Fraccaro, Denys God- win, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Gabby Nyirjesy, Blair Edwards, et al

    Johannes Jakubik, Sujit Roy, Christopher E. Phillips, Paolo Fraccaro, Denys God- win, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Gabby Nyirjesy, Blair Edwards, et al. 2023. Foundation Models for Generalist Geospatial Artificial Intelligence. arXiv:2310.18660 [cs.CV] https://arxiv.org/abs/2310.18660

    Pith/arXiv arXiv 2023

  14. [14]

    Simon Kornblith, Jonathon Shlens, and Quoc V. Le. 2019. Do Better ImageNet Models Transfer Better?. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2661–2671

    work page 2019

  15. [15]

    Alexandre Lacoste, Nils Lehmann, Pau Rodriguez, Evan Sherwin, Hannah Kerner, Bj"orn L"utjens, Jeremy Irvin, David Dao, Hamed Alemohammad, Alexandre Drouin, Mehmet Gunturkun, Gabriel Huang, David Vazquez, Dava Newman, Yoshua Bengio, Stefano Ermon, and Xiaoxiang Zhu. 2023. GEO-Bench: Toward Foundation Models for Earth Monitoring. InAdvances in Neural Inform...

    work page 2023

  16. [16]

    Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn

    Yoonho Lee, Annie S. Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn. 2023. Surgical Fine-Tuning Improves Adaptation to Distribution Shifts. InInternational Conference on Learning Representations

    work page 2023

  17. [17]

    Shikun Liu, Edward Johns, and Andrew J Davison. 2019. End-to-end multi-task learning with attention. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1871–1880

    work page 2019

  18. [18]

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. 2015. Fully Convolutional Networks for Semantic Segmentation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 3431–3440

    work page 2015

  19. [19]

    Gengchen Mai, Weiming Huang, Jin Sun, Shenzhe Song, Deepak Mishra, Ninghao Liu, Song Gao, Tianming Liu, Gao Cong, Yingjie Hu, Chris Cundy, Ziyuan Li, Rui Zhu, and Ni Lao. 2024. On the Opportunities and Challenges of Foundation Models for GeoAI.ACM Transactions on Spatial Algorithms and Systems10, 2, Article 11 (2024), 46 pages. doi:10.1145/3653070

    doi:10.1145/3653070 2024

  20. [20]

    Valerio Marsocci, Yuru Jia, Georges Le Bellier, David Kerekes, Liang Zeng, Se- bastian Hafner, Sebastian Gerard, Eric Brune, Ritu Yadav, Ali Shibli, Heng Fang, Yifang Ban, Maarten Vergauwen, Nicolas Audebert, and Andrea Nascetti. 2026. PANGAEA: Assessing Geospatial Foundation Models Capabilities through a Global and Inclusive Benchmark.IEEE Geoscience and...

    work page arXiv 2026

  21. [21]

    Francesc Marti Escofet, Benedikt Blumenstiel, Linus Scheibenreif, Paolo Fraccaro, and Konrad Schindler. 2025. Fine-tune smarter, not harder: Parameter-efficient fine-tuning for geospatial foundation models. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 516–532

    work page 2025

  22. [22]

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. 2016. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In International Conference on 3D Vision. 565–571

    work page 2016

  23. [23]

    Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. 2020. LEEP: A New Measure to Evaluate Transferability of Learned Representations. In Proceedings of the 37th International Conference on Machine Learning (Proceedings of Machine Learning Research)

    work page 2020

  24. [24]

    Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uyttendaele, and Trevor Darrell

    Colorado J. Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uyttendaele, and Trevor Darrell. 2023. Scale-MAE: A Scale-Aware Masked Autoencoder for Multiscale Geospatial Representation Learning. InProceedings of the IEEE/CVF International Conference on Computer Vision. 4088–4099

    work page 2023

  25. [25]

    Marina Sokolova and Guy Lapalme. 2009. A systematic analysis of performance measures for classification tasks.Information Processing & Management45, 4 (2009), 427–437

    work page 2009

  26. [26]

    Daniela Szwarcman, Sujit Roy, Paolo Fraccaro, Orsteinn Eli Gislason, Benedikt Blumenstiel, Rinki Ghosal, Pedro Henrique de Oliveira, Joao Lucas de Sousa Almeida, Rocco Sedona, Yanghui Kang, Srija Chakraborty, Sizhe Wang, Carlos Gomes, Ankur Kumar, Vishal Gaur, Myscon Truong, Denys Godwin, Sam Khallaghi, Hyunho Lee, Chia Yu Hsu, Ata Akbari Asanjan, Besart ...

    arXiv 2025

  27. [27]

    1979.Information Retrieval(2 ed.)

    Cornelis Joost van Rijsbergen. 1979.Information Retrieval(2 ed.). Butterworth- Heinemann

    work page 1979

  28. [28]

    Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. 2018. Unified Perceptual Parsing for Scene Understanding. InProceedings of the European Conference on Computer Vision. 418–434

    work page 2018

  29. [29]

    Stewart, Jo"elle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu

    Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J. Stewart, Jo"elle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. 2024. Neural Plasticity-Inspired Multimodal Foundation Model for Earth Observation. arXiv:2403.15356 [cs.CV] https://arxiv.org/abs/2403.15356

    arXiv 2024

  30. [30]

    Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. 2021. LogME: Practical Assessment of Pre-trained Models for Transfer Learning. InProceedings of the 38th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 139). PMLR, 12133–12143. https://proceedings.mlr.press/ v139/you21b.html SIGSPATIAL ’26, November 3–6,...

    work page 2021

  31. [2022]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Masked Autoencoders Are Scalable Vision Learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 16000–16009

    work page

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.