Pith. sign in

REVIEW 4 major objections 5 minor 28 references

Resampling Augmentation for Time Series Contrastive Learning: Application to Remote Sensing

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that resampling, not masking or jittering, is the best contrastive augmentation for satellite time series, and that it yields state-of-the-art S2-Agri100 accuracy without spatial or temporal encodings.

desk verdict A solid, well-tested resampling augmentation for satellite time series that likely beats common SSL baselines, but the 'consistently outperforms' and SOTA claims need baseline hyperparameter reporting, error bars, and a cleaner acknowledgment of the test-split pretraining before the strongest wording holds. read the letter →

arxiv 2506.18587 v1 pith:5OLEZYEN submitted 2025-06-23 cs.CV

Antoine Saget , Baptiste Lafabregue , Antoine Cornuéjols , Pierre Gançarski This is my paper · ORCID
classification cs.CV
keywords contrastivelearningself-supervisedtimeseriesaugmentationresamplingsatelliteimageSentinel-2cropclassificationlabelefficiency
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 is trying to establish that the choice of augmentation, not the architecture, is what limits contrastive self-supervised learning on satellite image time series. It introduces a resampling augmentation that upsamples each series, extracts two non-overlapping subsequences that each still cover the full temporal range, and interpolates them back to the original grid to form positive pairs. If the claim is right, then a cheap preprocessing step can replace more complex masking-based self-supervision and substantially reduce the labels needed for crop-type mapping. The reported experiments show improvements of up to 23 percentage points over raw features at five labels per class, and state-of-the-art performance on S2-Agri100 while using no spatial context and no temporal positional encoding.

What carries the argument

The load-bearing object is the resampling augmentation itself: a three-step transform with $T_{up}=2T$, two disjoint subsequences of length $T_{int}=T/2$, and a constraint that each quarter of the upsampled series contributes at least $\lfloor T_{int}/4\rfloor$ timesteps to each subsequence. The subsequences are then rescaled to the original temporal range and linearly interpolated back to the original timestamps. It is this quarter-coverage constraint that forces both views to preserve temporal coverage and prevents either view from dropping an entire phenological phase, while the upsampling gives the disjoint draws enough material to differ in phase.

What would settle it

Take a synthetic time series benchmark in which the class-defining event lasts only one or two timesteps and the sampling interval is comparable to the event length. If resampling with $T_{int}=T/2$ destroys that event in one of the two views, the positive pair will disagree on the label-relevant feature; showing that resampling then underperforms masking or raw features on this benchmark would falsify the claim that resampling is generally the best augmentation.

Watch

Extended reading notes

Core claim

The central discovery is that positive pairs for contrastive learning on time series can be generated by a resampling operation rather than by corrupting the signal. Starting from a series of $T$ timesteps, the method interpolates it to $2T$ timesteps, draws two disjoint subsequences of length $T/2$ under a quarter-coverage constraint, and then linearly rescales and interpolates each subsequence back to the original $T$ timesteps. The two views therefore share the same length, alignment, and overall phenological coverage while sampling the signal on different temporal grids. The paper argues that this makes the views 'hard' in the right way for the contrastive loss, and reports consistent gains over jittering, resizing, and masking across three datasets, a 4-10x reduction in labeled data needed to reach a given accuracy, and competitive or superior transfer to a geographically distinct benchmark.

Load-bearing premise

The method assumes that the time series is sampled densely enough that discarding half the timesteps (through subsampling) still preserves every event that matters for the downstream task; the paper explicitly says this holds for Sentinel-2's five-day revisit over crop cycles but can fail when meaningful events are rare or high-frequency.

Editorial extensions

If this is right

  • On FranceCrops, resampling lifts logistic-regression accuracy from 44% with raw features to 67% with only 5 labeled samples per class, implying a 4-10x reduction in required labeled data.
  • Resampling consistently beats jittering, resizing, and masking across all label budgets (5 to 100 samples per class) and on all three datasets, under both SimCLR and VICReg.
  • After pretraining on SITS-Former (California) and finetuning on S2-Agri100 (France), the method reaches 70.15% overall accuracy, above the SITS-Former and Presto baselines, despite ignoring spatial information and temporal positional encodings.
  • Pretraining on unlabeled S2-Agri100 data from the same domain raises finetuned accuracy to 76.84% and linear-evaluation accuracy to 74.30%, showing that domain-matched unlabeled data can be more valuable than a more complex classifier.
  • The small gap between linear evaluation and finetuning suggests the pretrained features carry most of the task information, so simpler evaluation protocols are sufficient for downstream use.

Reading between the lines

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

  • I would expect the quarter-coverage constraint to be the decisive design choice: without it, disjoint subsequences could each miss a transient event and the positive pairs would disagree on label-relevant structure. The paper does not report that ablation, but it is the natural next test.
  • The same mechanism should transfer to other quasi-periodic signals, such as electrocardiograms, weather series, or audio, provided the sampling rate exceeds the frequency of the events that define the downstream classes; this follows from the paper's own stated limitation rather than from its experiments.
  • An untested extension is to make the upsampling factor and subsequence length adaptive to irregularly sampled series or variable-length sequences, which would let the augmentation operate in a foundation-model-style setting instead of the fixed short-segment setting used here.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes a resampling-based augmentation for contrastive self-supervised learning on satellite image time series (SITS). Given an input series of length T, the method linearly upsamples it to T_up=2T, samples two disjoint subsequences of length T_int=T/2 under a per-quarter coverage constraint, and then linearly resamples each subsequence back to length T to form positive pairs. The authors evaluate the augmentation across four contrastive frameworks (SimCLR, BYOL, VICReg, MoCo) and three agricultural benchmarks (FranceCrops, FranceCrops Centre-Val de Loire, PASTIS), reporting logistic-regression label-efficiency curves from 5 to 100 labels per class and finetuning results on S2-Agri100. They report that resampling outperforms jittering, resizing, and masking across all sample sizes and datasets, achieves 70.15% overall accuracy on S2-Agri100 after pretraining on SITS-Former (versus 67.03% for SITS-Former and 68.89% for Presto), and that pretraining on target-domain S2-Agri100 data further raises accuracy to 76.84%. Code for models, training, and preprocessing is publicly released.

Significance. If the empirical claims hold, the paper makes a useful and simple contribution: a fixed, non-learned augmentation with two natural hyperparameters that improves label efficiency in SITS contrastive learning. The study's strengths include multi-framework and multi-dataset experiments with 20 runs and reported standard deviations, explicit non-optimization of the proposed hyperparameters, and public code. The paper also articulates a concrete limitation (dependence on a high temporal sampling rate relative to event frequency) that is falsifiable. The main threats are comparative rather than theoretical: the baseline augmentations are not calibrated or fully specified, the S2-Agri100 state-of-the-art claim rests on a small set of baselines with three runs and no variance reporting, and the target-domain pretraining experiment conflates data distribution with method.

major comments (4)
  1. [§4.3 / Table 3] The central claim of Section 5.2 that resampling 'consistently outperforms all other approaches across all sample sizes and datasets' is not fully supported because the experimental setup does not specify the hyperparameters of the baseline augmentations: there is no jittering amplitude, masking ratio or mask length, resizing factor, or statement of whether any of these values were tuned. Since Table 3 reports margins as small as one percentage point on PASTIS (e.g., 38 vs 37 at 5 samples per class), a reasonable strength sweep of the baselines could change the ranking. Please report the exact baseline settings used or, ideally, include a small strength sweep or point to the released configuration files so that the comparison is calibrated.
  2. [Tables 4–6] The state-of-the-art claim in the abstract and Section 5.3 is based on averages over only three runs with no standard deviations, and the comparison includes only SITS-Former and Presto. This is insufficient to support a literal 'state-of-the-art' claim on S2-Agri100. Please report per-run variability or confidence intervals, add at least one or two additional published baselines from the masked-modeling line, and either ensure the finetuning protocol is matched across methods or explicitly reposition the claim as 'outperforms the two compared baselines.'
  3. [§5.4 / Table 6] The comparison in Table 6 confounds augmentation and architecture with pretraining data: the rows using S2-Agri100 pretraining are not compared against SITS-Former or Presto under the same target-domain pretraining, so the 6.7-point gain over Table 4 cannot be attributed to the resampling augmentation or the model. The authors acknowledge this point in the text, but the abstract's state-of-the-art wording is still misleading. Please either run the comparison baselines under target-domain pretraining or clearly label the last two rows as 'resampling + target-domain pretraining' and remove the unqualified SOTA claim.
  4. [Table 3] Several of the headline improvements over masking are within the reported variability: all standard deviations are reported as ≤ 1 point, and on PASTIS the differences at 5, 10, and 20 labels are only 1 point (38 vs 37, 42 vs 41, 46 vs 45). Given that the 'consistently outperforms' claim is quantitative, please add a paired significance test or report, for each setting, the number of runs in which resampling exceeded each baseline, so that a systematic advantage can be distinguished from run-to-run noise.
minor comments (5)
  1. [Figure 1 caption] The caption contains 'Original time serie' and 'Subsampled time series'; the first should be 'Original time series' and the terminology should be consistent throughout.
  2. [Table 1 header] The header says 'supervsied downstream task evaluation'; this should be 'supervised downstream task evaluation'.
  3. [References] The Reed et al. reference contains 'UyttenDAele' with inconsistent capitalization; please fix.
  4. [§5.1] The sentence claiming that representation collapse occurs more frequently 'due to the lower dimensionality of time series data' is speculative and is not directly supported by the reported experiments; please soften it or provide evidence.
  5. [Abstract and §5.3] The phrase 'state-of-the-art' should be qualified as 'state-of-the-art among the methods compared here' unless a broader benchmark comparison is added.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the resampling augmentation is a fixed, non-learned transformation evaluated against external baselines and public benchmarks.

full rationale

No circular step is present in the paper. The resampling augmentation is defined explicitly in Section 3 (Eqs. 1-3) as a deterministic three-step transformation: linear upsampling, selection of two disjoint subsequences with per-quarter coverage constraints, and linear resampling back to the original temporal resolution. It contains no fitted parameters, no learned components, and no constants calibrated to the downstream labels or to the contrastive loss. The two hyperparameters T_up and T_int are set to natural values (T_up = 2T, T_int = T/2) and the paper explicitly states they were not optimized. Thus the method cannot reduce to its own inputs by construction. The paper's self-citations are confined to supporting infrastructure: the FranceCrops dataset and the multi-time-series aggregation module from Saget et al. (2024). These are used for evaluation and architectural plumbing, not as evidence for the central claim that resampling outperforms jittering, resizing, and masking; that claim is tested empirically against public benchmarks (PASTIS, S2-Agri100) and against standard contrastive frameworks (SimCLR, MoCo, BYOL, VICReg). There is no invocation of an author-imported uniqueness theorem, no ansatz smuggled in through citation, and no renaming of a known result in new coordinates. The acknowledged limitation in Section 6, namely that the method requires a high temporal sampling rate relative to meaningful event frequency, is an honest scope condition and not a circular dependency. Even if baseline augmentation strengths are under-specified, that is an empirical calibration concern about the strength of the comparison, not a circularity of derivation, and the instruction here is to reserve circularity findings for cases where a paper's own equations or fitted parameters force the result. No such case arises.

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

The ledger shows no invented entities or fitted constants. The two augmentation hyperparameters are set to natural values and explicitly not optimized; the main experimental choices, such as G, batch size, and optimizer, are conventional or from prior work.

free parameters (3)
  • T_up (upsampled sequence length) = 2 x T
    Set to twice the original length; stated as a natural value, not optimized. Central to the augmentation's temporal resolution.
  • T_int (subsequence length) = T / 2
    Set to half the original length; stated as a natural value, not optimized. Controls how much temporal structure is removed.
  • G (number of sampled pixel time series per sample) = 4
    Borrowed from Saget et al. 2024; used for aggregating multiple pixel time series. Not part of the augmentation but affects all experiments.
assumptions (3)
  • domain assumption Linear interpolation adequately represents the satellite time series between observed timesteps.
    Used in Equations (1) and (3) to upsample and resample views; if false, the views introduce interpolation artifacts.
  • domain assumption Temporal sampling rate is high enough that subsampling to half length preserves discriminative events.
    Explicitly stated as a limitation in Section 6; load-bearing for the method's validity on the tested data.
  • domain assumption Contrastive self-supervised pretraining on unlabeled SITS transfers to downstream crop classification.
    Standard SSL assumption underlying all experiments and the label-efficiency claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Resampling Augmentation for Time Series Contrastive Learning: Application to Remote Sensing." pith.science (2026). https://pith.science/paper/5OLEZYEN

@misc{pith2026250618587,
  author       = {Pith},
  title        = {Pith review of: Resampling Augmentation for Time Series Contrastive Learning: Application to Remote Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5OLEZYEN}},
  note         = {Machine review of arXiv:2506.18587}
}
read the original abstract

Given the abundance of unlabeled Satellite Image Time Series (SITS) and the scarcity of labeled data, contrastive self-supervised pretraining emerges as a natural tool to leverage this vast quantity of unlabeled data. However, designing effective data augmentations for contrastive learning remains challenging for time series. We introduce a novel resampling-based augmentation strategy that generates positive pairs by upsampling time series and extracting disjoint subsequences while preserving temporal coverage. We validate our approach on multiple agricultural classification benchmarks using Sentinel-2 imagery, showing that it outperforms common alternatives such as jittering, resizing, and masking. Further, we achieve state-of-the-art performance on the S2-Agri100 dataset without employing spatial information or temporal encodings, surpassing more complex masked-based SSL frameworks. Our method offers a simple, yet effective, contrastive learning augmentation for remote sensing time series.

Figures

Figures reproduced from arXiv: 2506.18587 by the authors.

Figure 1
Figure 1. Example visualization of the resampling augmen [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Self-Supervised Contrastive retraining. For each sample [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 10 canonical work pages

  1. [1]

    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 format.date year duplicate empty "emp...

  2. [2]

    Vicreg: Variance-invariance-covariance regularization for self-supervised learning

    Bardes, A., Ponce, J., and LeCun, Y. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021

    arXiv 2021

  3. [3]

    Unsupervised learning of visual features by contrasting cluster assignments

    Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., and Joulin, A. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33: 0 9912--9924, 2020

    2020

  4. [4]

    A simple framework for contrastive learning of visual representations

    Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pp.\ 1597--1607. PMLR, 2020 a

    2020

  5. [5]

    Chen, T., Kornblith, S., Swersky, K., Norouzi, M., and Hinton, G. E. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33: 0 22243--22255, 2020 b

    2020

  6. [6]

    and He, K

    Chen, X. and He, K. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 15750--15758, 2021

    2021

  7. [7]

    Timemae: Self-supervised representations of time series with decoupled masked autoencoders

    Cheng, M., Liu, Q., Liu, Z., Zhang, H., Zhang, R., and Chen, E. Timemae: Self-supervised representations of time series with decoupled masked autoencoders. arXiv preprint arXiv:2303.00320, 2023

    arXiv 2023

  8. [8]

    Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery

    Cong, Y., Khanna, S., Meng, C., Liu, P., Rozi, E., He, Y., Burke, M., Lobell, D., and Ermon, S. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35: 0 197--211, 2022

    work page 2022

Show all 28 references
  1. [9]

    Sentinel-2: Esa's optical high-resolution mission for gmes operational services

    Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B., Isola, C., Laberinti, P., Martimort, P., et al. Sentinel-2: Esa's optical high-resolution mission for gmes operational services. Remote sensing of Environment, 120: 0 25--36, 2012

    2012

  2. [10]

    Garnot, V. S. F. and Landrieu, L. Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 4872--4881, 2021

    2021

  3. [11]

    Garnot, V. S. F., Landrieu, L., Giordano, S., and Chehata, N. Satellite image time series classification with pixel-set encoders and temporal self-attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12325--12334, 2020

    2020

  4. [12]

    Copernicus sentinel-2a calibration and products validation status

    Gascon, F., Bouzinac, C., Th \'e paut, O., Jung, M., Francesconi, B., Louis, J., Lonjou, V., Lafrance, B., Massera, S., Gaudel-Vacaresse, A., et al. Copernicus sentinel-2a calibration and products validation status. Remote Sensing, 9 0 (6): 0 584, 2017

    2017

  5. [13]

    Bootstrap your own latent-a new approach to self-supervised learning

    Grill, J.-B., Strub, F., Altch \'e , F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33: ...

    2020

  6. [14]

    Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery

    Guo, X., Lao, J., Dang, B., Zhang, Y., Yu, L., Ru, L., Zhong, L., Huang, Z., Wu, K., Hu, D., et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vis...

    2024

  7. [15]

    Momentum contrast for unsupervised visual representation learning

    He, K., Fan, H., Wu, Y., Xie, S., and Girshick, R. Momentum contrast for unsupervised visual representation learning. arxiv e-prints, art. arXiv preprint arXiv:1911.05722, 2019

    1911 arXiv

  8. [16]

    Masked autoencoders are scalable vision learners

    He, K., Chen, X., Xie, S., Li, Y., Doll \'a r, P., and Girshick, R. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 16000--16009, 2022

    2022

  9. [17]

    Data-efficient image recognition with contrastive predictive coding

    Henaff, O. Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning, pp.\ 4182--4192. PMLR, 2020

    2020

  10. [18]

    Foundation models for generalist geospatial artificial intelligence

    Jakubik, J., Roy, S., Phillips, C., Fraccaro, P., Godwin, D., Zadrozny, B., Szwarcman, D., Gomes, C., Nyirjesy, G., Edwards, B., et al. Foundation models for generalist geospatial artificial intelligence. CoRR, 2023

    2023

  11. [19]

    Guidelines for augmentation selection in contrastive learning for time series classification

    Liu, Z., Alavi, A., Li, M., and Zhang, X. Guidelines for augmentation selection in contrastive learning for time series classification. arXiv preprint arXiv:2407.09336, 2024

    2024 arXiv

  12. [20]

    Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data

    Manas, O., Lacoste, A., Gir \'o -i Nieto, X., Vazquez, D., and Rodriguez, P. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 9414--9423, 2021

    2021

  13. [21]

    Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning

    Reed, C., Gupta, R., Li, S., Brockman, S., Funk, C., Clipp, B., Candido, S., UyttenDAele, M., and Darrell, T. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. 2023 ieee. In CVF International Conference on Computer Vision (ICCV), pp...

    2023

  14. [22]

    Learning from few labeled time series with segment-based self-supervised learning: application to remote-sensing

    Saget, A., Lafabregue, B., Cornu \'e jols, A., and Gan c arski, P. Learning from few labeled time series with segment-based self-supervised learning: application to remote-sensing. In Proceedings of SPAICE2024: The First Joint European Space Agency/IAA Conference on AI in and ...

    2024

  15. [23]

    Lightweight, pre-trained transformers for remote sensing timeseries

    Tseng, G., Cartuyvels, R., Zvonkov, I., Purohit, M., Rolnick, D., and Kerner, H. Lightweight, pre-trained transformers for remote sensing timeseries. arXiv preprint arXiv:2304.14065, 2023

    2023 arXiv

  16. [24]

    Ssl4eo-s12: a large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation

    Wang, Y., Braham, N., Xiong, Z., Liu, C., Albrecht, C., Zhu, X., et al. Ssl4eo-s12: a large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation. arxiv. arXiv preprint arXiv:2211.07044, 10, 2022

    2022 arXiv

  17. [25]

    M., Braham, N

    Wang, Y., Albrecht, C. M., Braham, N. A. A., Liu, C., Xiong, Z., and Zhu, X. X. Decur: decoupling common & unique representations for multimodal self-supervision. arXiv preprint arXiv:2309.05300, 2023

    2023 arXiv

  18. [26]

    Time series classification from scratch with deep neural networks: a strong baseline

    Wang, Z., Yan, W., and Oates, T. Time series classification from scratch with deep neural networks: a strong baseline. corr abs/1611.06455 (2016). arXiv preprint arXiv:1611.06455, 2016

    2016 arXiv

  19. [27]

    Sits-former: A pre-trained spatio-spectral-temporal representation model for sentinel-2 time series classification

    Yuan, Y., Lin, L., Liu, Q., Hang, R., and Zhou, Z.-G. Sits-former: A pre-trained spatio-spectral-temporal representation model for sentinel-2 time series classification. International Journal of Applied Earth Observation and Geoinformation, 106: 0 102651, 2022

    2022

  20. [28]

    Barlow twins: Self-supervised learning via redundancy reduction

    Zbontar, J., Jing, L., Misra, I., LeCun, Y., and Deny, S. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pp.\ 12310--12320. PMLR, 2021

    2021

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.