Pith. sign in

REVIEW 4 major objections 5 minor 61 references

A Joint Learning Framework with Feature Reconstruction and Prediction for Incomplete Satellite Image Time Series in Agricultural Semantic Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a joint training framework—reconstructing only the temporal features a segmentation model needs, guided by a frozen teacher trained on complete time series—lets agricultural segmentation models handle cloud-induced…

desk verdict Solid incremental framework for incomplete SITS segmentation; gains are plausible but the single-run margins need error bars before the headline holds. read the letter →

arxiv 2505.19159 v1 pith:4U7JAFU6 submitted 2025-05-25 cs.CV

Yuze Wang , Mariana Belgiu , Haiyang Wu , Dandan Zhong , Yangyang Cao , Chao Tao This is my paper · ORCID
classification cs.CV
keywords SatelliteimagetimeseriesincompleteSITSfeaturereconstructionknowledgedistillationtemporalmaskingagriculturalsemanticsegmentationcropclassificationcroplandextraction
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

Clouds create gaps in satellite image time series that break temporal dependencies and shift feature distributions, so models trained on complete sequences lose accuracy. The paper tries to establish that a model can be trained to handle such gaps without reconstructing the whole sequence and without learning shortcuts: during training, a student segmentation model sees randomly masked time series, reconstructs only the deepest temporal features that a frozen teacher (trained on complete series) would have produced, and is simultaneously supervised by ground-truth labels and teacher logits. The authors argue this makes reconstruction selective, prevents noise from full imputation, and stops augmented models from overfitting to missing patterns. Across three study regions and two sensors, the method improves mean F1 by 6.93% for cropland extraction and 7.09% for crop classification over compared methods, and it transfers across seven backbones. If correct, this offers a model-agnostic recipe for making agricultural monitoring robust to cloudy periods.

What carries the argument

The load-bearing object is the feature-reconstruction loss applied to the deepest temporally fused feature, standardized channel-wise, with the frozen teacher as target. For each backbone this feature is named differently—U-TAE's temporal-attention-fused encoder output, 3D CNN's last encoder layer, RNN hidden states, or TSViT's class-token embedding—which is why the framework is backbone-agnostic. The standardization $Z = (F - \mu_F)/\sigma_F$ along the channel dimension converts the teacher-student comparison from absolute magnitudes to distributional patterns, removing the global feature shift caused by missing time steps. The prediction losses then ensure that the reconstructed features are the ones that actually help classification, so the student does not waste capacity on irrelevant detail and does not fall into shortcut reasoning that skips long temporal dependencies.

What would settle it

On an independent multi-year SITS crop-classification benchmark, apply the same 25%-75% temporal masking protocol and compare mean F1 with DA-WS and DA-TD; if the joint framework does not beat the best augmentation baseline, the claimed margin is dataset-specific rather than a general property of the method.

Watch

Extended reading notes

Core claim

The central claim is that incomplete satellite image time series should be handled at the feature level, not the pixel or data level. The proposed joint learning framework trains a student model on randomly masked inputs $X'$ and uses a frozen teacher $f_T$ pre-trained on complete SITS to supervise two coupled goals: a feature reconstruction loss $L_{\mathrm{FR}} = \frac{1}{N}\sum_i \|Z_i - Z'_i\|_2^2$ between standardized deepest temporal features $Z$ and $Z'$, plus prediction losses (cross-entropy with labels and KL divergence between student and teacher logits). The authors report mean F1 improvements of up to 6.93% in cropland extraction and 7.09% in crop classification over data-reconstruction and data-augmentation baselines, with the largest gains in real-world continuous-gap settings. They also state that the framework retains performance on complete SITS while improving accuracy under 25% to 75% temporal missing rates, across Sentinel-2 and PlanetScope data and across RNN, 3D-CNN, and transformer backbones.

Load-bearing premise

The framework assumes that a teacher model trained only on complete time series gives trustworthy temporal features for every class; the authors themselves report that for leguminous crops, where within-class growth cycles vary widely, the teacher's errors are passed to the student.

Editorial extensions

If this is right

  • Models trained with this framework maintain near-complete accuracy on full time series while gaining accuracy at 25%, 50%, and 75% temporal missing rates, so it can be deployed without sacrificing the complete-data regime.
  • The framework is backbone-agnostic: improvements hold across ConvLSTM, ConvGRU, U-ConvLSTM, FPN-ConvLSTM, 3D U-Net, U-TAE, and TSViT, with the largest relative gains on the backbones that fail most on incomplete inputs.
  • Because it works directly on masked SITS without external SAR data or cloud-free reference imagery, it lowers the cost of operational agricultural monitoring in persistently cloudy regions.
  • The method transfers across Sentinel-2 and PlanetScope, so its benefit does not depend on one sensor's temporal resolution.

Reading between the lines

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

  • Because the teacher is frozen and trained only on complete SITS, the framework's ceiling is bounded by how well complete-data models can represent phenology; the authors' own leguminous-crop errors suggest that classes with high intra-class variability would need a stronger or multi-teacher target.
  • The channel-wise standardization implies the loss is invariant to per-channel affine shifts, so a natural testable extension is to replace z-score normalization with a learned affine-invariant distance or to apply the same idea to spatial features, which the paper explicitly leaves out.
  • For short-cycle crops such as vegetables and melons, the method fails when masking removes the entire growth window; combining temporal masking with class-aware sampling of windows might preserve anchor points for those crops.
  • The gains over temporal dropout at 0% missingness suggest self-distillation acts as a regularizer, so a plausible extension is to test whether the framework also reduces shortcut reliance under other distribution shifts, such as cross-year or cross-region transfer.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes a joint learning framework for semantic segmentation of satellite image time series with missing temporal observations. A teacher model is pretrained on complete SITS; a student model is trained on temporally masked inputs under three losses: feature reconstruction toward the teacher's standardized temporally fused features, logit distillation, and ground-truth cross-entropy. The method is evaluated on cropland extraction (Hunan SEN and Hunan PLA, Sentinel-2 and PlanetScope) and crop classification (Fr&Cat S4A), in simulated and real-world incomplete settings, across seven backbones, with ablations of the mask, distillation, feature-reconstruction, and standardization modules. The paper reports consistent improvements over data-reconstruction and data-augmentation baselines and releases code.

Significance. If the reported gains are reproducible across seeds, the framework is a useful model-agnostic training strategy for incomplete SITS, with unusually broad empirical coverage: three datasets, two sensors, two tasks, simulated and real-world missingness, and seven backbones. The ablation isolating the standardization step is a nice mechanistic insight, and the public code will help adoption. The main qualification is that the central comparative claim currently rests on single-run experiments and on averaged relative headline numbers.

major comments (4)
  1. [§4.1, Tables 1–4] All reported results are single runs with no seed information, standard deviations, or significance tests. The mean-F1 margins over the best data-augmentation baseline are small in the cropland experiments: +1.79 (Table 1), +1.39 (Table 2), +1.23 (Table 3), and +1.33 (Table 4). Since seed-to-seed variation of U-TAE-based models in agricultural segmentation is commonly on the order of 1–2 mean-F1 points, these margins do not by themselves establish that the proposed method outperforms DA baselines. I request multi-seed runs (at least five seeds) with mean±std, or paired significance tests, for the central comparisons; the crop-classification margins are larger, but the cropland claim specifically depends on this evidence.
  2. [Abstract, §4.2] The abstract's headline gains, '6.93% in cropland extraction and 7.09% in crop classification', are not values found in any single table. They equal the average of relative (percentage-of-baseline) improvements over the unaugmented Baseline across the two settings per task: (2.89+18.04+4.32+2.48)/4 = 6.93 and (10.35+3.83)/2 = 7.09 for Tables 1–4 and Tables 5 and 7, respectively. The 18.04% term comes from a setting where the baseline is unusually low (69.46 M-F1), so the average overstates typical gains. The paper should report absolute per-setting margins and define the headline statistic explicitly.
  3. [§3.2, Eq. (6)] The loss weights σ, γ, λ, the distillation temperature T, and the mask-ratio range M,N are not specified in Section 4.1; the sentence defining M,N is incomplete ('typically set to 25...'). These hyperparameters control the balance between reconstruction and prediction and the simulated missingness distribution, so they are needed to reproduce the method from the paper; a sensitivity analysis would also strengthen the claim that the framework is robust to their choice.
  4. [§4.3, Tables 11–12] On the Fr&Cat S4A crop-classification task at 0% missingness, the proposed method's mean F1 (84.04) is below the baseline (84.77), and the text's summary 'it still keeps a mean F1-scores of 80.47% under complete SITS, low (25%), and middle (50%) level missing conditions' conflates the 50% result with the complete/low conditions. This does not invalidate the method, but the claim that it 'maintains strong classification capability on complete time series' is only supported for cropland extraction, not for crop classification; the paper should reconcile this nuance with the stated robustness claim.
minor comments (5)
  1. [Throughout] There are several typos and formatting errors: 'U-TILIES' in Tables 1 and 2 should be 'U-TILISE'; 'gdecreases' in Section 4.2.2; 'loss its ability' in Appendix A; 'leanring' in the Introduction; 'One the one hand' in Section 3.2; and 'STIS' in Section 2.1.
  2. [§4.3] Reference [51] is cited for the self-distillation process, but [51] is the TSViT architecture paper; the self-distillation regularization source appears to be [52] (Mobahi et al.), so the citation should be corrected.
  3. [§3.2, Eq. (2)] Please clarify over which dimensions the mean and standard deviation in Eq. (2) are computed: per-sample over the channel dimension only, or over spatial and temporal axes as well. The current description ('along the channel dimensional') is ambiguous and matters for reproducibility.
  4. [Tables 15–16] The header row with the module indicators MA, KD, FR, FS is hard to parse; please reformat the module-combination rows so that each row explicitly names the active modules (e.g., 'MA+KD+FR') and the check/cross markers align with the columns.
  5. [Figure 4] The toy case in Figure 4 would benefit from a formal statement of the claim and a description of how S1 and S2 are constructed; the caption is not self-contained as written.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: teacher-guided training targets are used only during training; test evaluations are on independent held-out splits.

full rationale

The paper's central claim is that joint feature-reconstruction and prediction training improves robustness to missing temporal observations. The derivation chain is not circular: the teacher model is pre-trained on the complete training subset and frozen (Sec. 3.2), while the student is trained on temporally masked versions of the same training data. The feature-reconstruction loss L_FR (Eq. 3) and logit-distillation loss L_KD (Eq. 5) are training objectives only; they are anchored to teacher outputs, but the teacher is not fitted to test data. All reported test results come from held-out splits: simulated experiments drop 25%–75% of time steps on the 40% test portion of the complete subsets, and real-world experiments test on the separate incomplete subsets (Sec. 4.1). Neither the teacher nor any loss is fit to test labels or test masks, so the improvements are not forced by construction. The self-citations (e.g., [3], [9], [17]) appear only as background references for deep-learning segmentation, temporal class separability, and cloud removal; none carries the load-bearing argument. The acknowledged teacher-quality limitation for LC/V&M classes (Sec. 4.3) is a performance caveat, not a circular dependency. Therefore no circular step is identified.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The framework's load-bearing assumptions are the reliability of the teacher model and the sufficiency of the chosen feature layer; these are empirical design choices rather than mathematically forced constraints.

free parameters (3)
  • loss weights sigma, gamma, lambda = not reported
    Coefficients balancing feature reconstruction, knowledge distillation, and cross-entropy losses in Eq. 6; chosen by validation but values are not given in the paper.
  • distillation temperature T = not reported
    Temperature in the KL divergence loss of Eq. 5; no value is provided.
  • temporal mask range M, N = 25% to 75% (inferred)
    The sentence defining M and N is cut off: 'typically set to 25'; the experiments drop 25%-75% of time steps, so the intended range is clear.
assumptions (4)
  • domain assumption The teacher model trained on complete SITS provides a reliable feature representation to supervise the student.
    The entire feature reconstruction loss L_FR in Eq. 3 assumes the teacher's features are a valid target; the paper acknowledges failures for LC class where teacher knowledge is poor.
  • domain assumption The deepest temporally fused feature is sufficient to represent task-relevant temporal dynamics.
    Section 3.2 chooses the deepest fused feature as the supervision target, discarding spatial and lower-level features.
  • domain assumption Random temporal masks with 25%-75% missing ratios during training cover the distribution of real-world gaps.
    The training simulates missingness with uniform random masks; real-world gaps are often continuous and seasonal, which the real-world experiments partly validate but do not fully cover.
  • standard math Standard stochastic gradient optimization and generalization assumptions for deep networks.
    The method relies on standard backpropagation and the assumption that training on a sample of sub-regions generalizes across regions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Joint Learning Framework with Feature Reconstruction and Prediction for Incomplete Satellite Image Time Series in Agricultural Semantic Segmentation." pith.science (2026). https://pith.science/paper/4U7JAFU6

@misc{pith2026250519159,
  author       = {Pith},
  title        = {Pith review of: A Joint Learning Framework with Feature Reconstruction and Prediction for Incomplete Satellite Image Time Series in Agricultural Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4U7JAFU6}},
  note         = {Machine review of arXiv:2505.19159}
}
read the original abstract

Satellite Image Time Series (SITS) is crucial for agricultural semantic segmentation. However, Cloud contamination introduces time gaps in SITS, disrupting temporal dependencies and causing feature shifts, leading to degraded performance of models trained on complete SITS. Existing methods typically address this by reconstructing the entire SITS before prediction or using data augmentation to simulate missing data. Yet, full reconstruction may introduce noise and redundancy, while the data-augmented model can only handle limited missing patterns, leading to poor generalization. We propose a joint learning framework with feature reconstruction and prediction to address incomplete SITS more effectively. During training, we simulate data-missing scenarios using temporal masks. The two tasks are guided by both ground-truth labels and the teacher model trained on complete SITS. The prediction task constrains the model from selectively reconstructing critical features from masked inputs that align with the teacher's temporal feature representations. It reduces unnecessary reconstruction and limits noise propagation. By integrating reconstructed features into the prediction task, the model avoids learning shortcuts and maintains its ability to handle varied missing patterns and complete SITS. Experiments on SITS from Hunan Province, Western France, and Catalonia show that our method improves mean F1-scores by 6.93% in cropland extraction and 7.09% in crop classification over baselines. It also generalizes well across satellite sensors, including Sentinel-2 and PlanetScope, under varying temporal missing rates and model backbones.

Figures

Figures reproduced from arXiv: 2505.19159 by the authors.

Figure 1
Figure 1. The location and extent of datasets, where red and blue regions represent [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Monthly data availability of SITS samples in incomplete subsets: Hunan SEN, [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. The general framework of the proposed method [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The toy case where two students, S1 and S2, are guided by the same teacher T. Before the standardization(Z-score), S1 reconstructed features from incomplete SITS that are more similar to the teacher’s ideal features, yet it achieves a lower F1-score. In contrast, featu…
Figure 5
Figure 5. Figure 5: Visualization of sample results for our and compared methods in the Hunan [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Visualization of sample results for our and compared methods in the Hunan [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: Visualization of sample results for our and compared methods in Fr&Cat S4A [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 33 canonical work pages

  1. [1]

    Zhang, Y

    D. Zhang, Y. Pan, J. Zhang, T. Hu, J. Zhao, N. Li, Q. Chen, A gen- eralized approach based on convolutional neural networks for large area cropland mapping at very high resolution, Remote Sensing of Environ- ment 247 (2020-09) 111912. doi:10.1016/j.rse.2020.111912

    work page arXiv 2020

  2. [2]

    L. Wang, J. Wang, Z. Liu, J. Zhu, F. Qin, Evaluation of a deep- learning model for multispectral remote sensing of land use and crop classification, The Crop Journal 10 (5) (2022-10-01) 1435–1451. doi: 10.1016/j.cj.2022.01.009

    work page doi:10.1016/j.cj.2022.01.009 2022

  3. [3]

    H. Wu, Z. Du, D. Zhong, Y. Wang, C. Tao, Fsvlm: A vision-language model for remote sensing farmland segmentation, IEEE Transactions on Geoscience and Remote Sensing (2025). doi:https://doi.org/10. 1109/TGRS.2025.3532960

    arXiv 2025

  4. [4]

    Jiang, X

    T. Jiang, X. Liu, L. Wu, Method for Mapping Rice Fields in Complex Landscape Areas Based on Pre-Trained Convolutional Neural Network from HJ-1 A/B Data, ISPRS International Journal of Geo-Information 7 (11) (2018-10-30) 418. doi:10.3390/ijgi7110418

    work page doi:10.3390/ijgi7110418 2018

  5. [5]

    Persello, V

    C. Persello, V. Tolpekin, J. Bergado, Delineation of agricultural fields in smallholder farms from satellite images using fully convolutional net- works and combinatorial grouping, Remote Sensing of Environment 231 (2019-09) 111253. 39

    work page 2019

  6. [6]

    Gebbers, V

    R. Gebbers, V. I. Adamchuk, Precision Agriculture and Food Secu- rity, Science 327 (5967) (2010-02-12) 828–831. doi:10.1126/science. 1183899

    doi:10.1126/science 2010

  7. [7]

    J. Xu, J. Yang, X. Xiong, H. Li, J. Huang, K. Ting, Y. Ying, T. Lin, Towards interpreting multi-temporal deep learning models in crop map- ping, Remote Sensing of Environment 264 (2021-10) 112599. doi: 10.1016/j.rse.2021.112599

    work page arXiv 2021

  8. [8]

    Waldner, G

    F. Waldner, G. S. Canto, P. Defourny, Automated annual cropland map- ping using knowledge-based temporal features, ISPRS Journal of Pho- togrammetry and Remote Sensing 110 (2015-12) 1–13. doi:10.1016/ j.isprsjprs.2015.09.013

    work page 2015

Show all 61 references
  1. [9]

    Belgiu, O

    M. Belgiu, O. Csillik, Sentinel-2 cropland mapping using pixel-based and object-based time-weighted dynamic time warping analysis, Remote Sensing of Environment 204 (2018-01) 509–523. doi:10.1016/j.rse. 2017.10.005

    2018 doi

  2. [10]

    Zhang, W

    Y. Zhang, W. B. Rossow, A. A. Lacis, V. Oinas, M. I. Mishchenko, Calculation of radiative fluxes from the surface to top of atmosphere based on ISCCP and other global data sets: Refinements of the radiative transfer model and the input data, Journal of Geophysical Research: At...

    2004 doi

  3. [11]

    K. Mao, Z. Yuan, Z. Zuo, T. Xu, X. Shen, C. Gao, Changes in global cloud cover based on remote sensing data from 2003 to 2012, Chinese Geographical Science 29 (2019) 306–315. doi:10.1007/ s11769-019-1030-6

    2019

  4. [12]

    Z. Gu, P. Ebel, Q. Yuan, M. Schmitt, X. X. Zhu, Explicit Haze & Cloud Removal for Global Land Cover Classification, in: CVPR 2022 Workshop on Multimodal Learning for Earth and Environment, 2022-07, pp. 1–6

    2022

  5. [13]

    J. Wu, T. Li, L. Lin, C. Zeng, Progressive gap-filling in optical remote sensing imagery through a cascade of temporal and spatial reconstruc- tion models, Remote Sensing of Environment 311 (2024-09-01) 114245. doi:10.1016/j.rse.2024.114245. 40

    2024

  6. [14]

    Z. Li, H. Shen, Q. Weng, Y. Zhang, P. Dou, L. Zhang, Cloud and cloud shadow detection for optical satellite imagery: Features, algorithms, val- idation, and prospects, ISPRS Journal of Photogrammetry and Remote Sensing 188 (2022-06) 89–108. doi:10.1016/j.isprsjprs.2022.03. 020

    2022 doi

  7. [15]

    O. G. Ajayi, A. Ojima, Performance evaluation of selected cloud oc- clusion removal algorithms on remote sensing imagery, Remote Sens- ing Applications: Society and Environment 25 (2022-01-01) 100700. doi:10.1016/j.rsase.2022.100700

    2022

  8. [16]

    Wang, X.-L

    J.-L. Wang, X.-L. Zhao, H.-C. Li, K.-X. Cao, J. Miao, T.-Z. Huang, Unsupervised Domain Factorization Network for Thick Cloud Removal of Multitemporal Remotely Sensed Images, IEEE Transactions on Geo- science and Remote Sensing 61 (2023) 1–12. doi:10.1109/TGRS.2023. 3303169

    2023 doi

  9. [17]

    C. Tao, S. Fu, J. Qi, H. Li, Thick Cloud Removal in Optical Remote Sensing Images Using a Texture Complexity Guided Self-Paced Learn- ing Method, IEEE Transactions on Geoscience and Remote Sensing 60 (2022) 1–12. doi:10.1109/TGRS.2022.3157917

    2022

  10. [18]

    H. Shen, X. Li, Q. Cheng, C. Zeng, G. Yang, H. Li, L. Zhang, Missing Information Reconstruction of Remote Sensing Data: A Technical Re- view, IEEE Geoscience and Remote Sensing Magazine 3 (3) (2015-09) 61–85. doi:10.1109/MGRS.2015.2441912

    2015

  11. [19]

    Pipia, J

    L. Pipia, J. Mu˜ noz-Mar ´ ı, E. Amin, S. Belda, G. Camps-Valls, J. Verrelst, Fusing optical and SAR time series for LAI gap filling with multioutput Gaussian processes, Remote Sensing of Environment 235 (2019-12-15) 111452. doi:10.1016/j.rse.2019.111452

    2019

  12. [20]

    Tsokas, M

    A. Tsokas, M. Rysz, P. M. Pardalos, K. Dipple, SAR data applications in earth observation: An overview, Expert Systems with Applications 205 (2022-11-01) 117342. doi:10.1016/j.eswa.2022.117342

    2022

  13. [21]

    Y. Yuan, L. Lin, Z.-G. Zhou, H. Jiang, Q. Liu, Bridging optical and SAR satellite image time series via contrastive feature extraction for crop classification, ISPRS Journal of Photogrammetry and Remote Sensing 195 (2023-01-01) 222–232. doi:10.1016/j.isprsjprs.2022.11.020. 41

    2023 doi

  14. [22]

    Lepot, J.-B

    M. Lepot, J.-B. Aubin, F. H. L. R. Clemens, Interpolation in Time Series: An Introductive Overview of Existing Methods, Their Perfor- mance Criteria and Uncertainty Assessment, Water 9 (10) (2017-10)

    2017

  15. [23]

    W. R. Moskola ¨ ı, W. Abdou, A. Dipanda, Kolyang, Application of Deep Learning Architectures for Satellite Image Time Series Predic- tion: A Review, Remote Sensing 13 (23) (2021-01) 4822. doi:10.3390/ rs13234822

    2021

  16. [24]

    L. Li, J. Yan, H. Wang, Y. Jin, Anomaly Detection of Time Series With Smoothness-Inducing Sequential Variational Auto-Encoder, IEEE Transactions on Neural Networks and Learning Systems 32 (3) (2021-03) 1177–1191. doi:10.1109/TNNLS.2020.2980749

    2021

  17. [25]

    W. Quan, J. Chen, Y. Liu, D.-M. Yan, P. Wonka, Deep Learning- Based Image and Video Inpainting: A Survey, International Journal of Computer Vision 132 (7) (2024-07-01) 2367–2400. doi:10.1007/ s11263-023-01977-6

    2024

  18. [26]

    C. Fang, C. Wang, Time Series Data Imputation: A Survey on Deep Learning Approaches (2020-11-23). arXiv:2011.11347, doi:10.48550/ arXiv.2011.11347

    2020 doi arXiv

  19. [27]

    Gangopadhyay, S

    T. Gangopadhyay, S. Y. Tan, Z. Jiang, R. Meng, S. Sarkar, Spatiotem- poral Attention for Multivariate Time Series Prediction and Interpreta- tion, in: ICASSP 2021 - 2021 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2021-06, pp. 3560–3564...

    2021

  20. [28]

    X. Cai, Y. Bi, P. Nicholl, R. Sterritt, Revisiting the Encoding of Satellite Image Time Series (2023-09-08). arXiv:2305.02086, doi:10.48550/ arXiv.2305.02086

    2023 doi arXiv

  21. [29]

    Casolaro, V

    A. Casolaro, V. Capone, G. Iannuzzo, F. Camastra, Deep Learning for Time Series Forecasting: Advances and Open Problems, Information 14 (11) (2023-11) 598. doi:10.3390/info14110598

    2023 doi

  22. [30]

    Q. Wen, L. Sun, F. Yang, X. Song, J. Gao, X. Wang, H. Xu, Time series data augmentation for deep learning: A survey, International Joint 42 Conferences on Artificial Intelligence Organization (2021) pages=306– 315,doi:https://doi.org/10.24963/ijcai.2021/631

    2021 doi

  23. [31]

    Z. Gao, H. Liu, L. Li, Data Augmentation for Time-Series Classification: An Extensive Empirical Study and Comprehensive Survey (2024-08-24). arXiv:2310.10060, doi:10.48550/arXiv.2310.10060

    2024 doi arXiv

  24. [32]

    A. L. Guennec, S. Malinowski, R. Tavenard, Data Augmentation for Time Series Classification using Convolutional Neural Networks, in: ECML/PKDD Workshop on Advanced Analytics and Learning on Tem- poral Data, 2016-09-19

    2016

  25. [33]

    Kirkpatrick, R

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran, R. Hadsell, Overcoming catas- trophic forgetting in neural networks, Proceedings of the National Academy o...

    2017

  26. [34]

    Schak, A

    M. Schak, A. Gepperth, A study on catastrophic forgetting in deep lstm networks, in: Artificial Neural Networks and Machine Learning–ICANN 2019: Deep Learning: 28th International Conference on Artificial Neural Networks, Munich, Germany, September 17–19, 2019, Proceedings, Par...

    2019

  27. [35]

    Geirhos, J.-H

    R. Geirhos, J.-H. Jacobsen, C. Michaelis, R. Zemel, W. Brendel, M. Bethge, F. A. Wichmann, Shortcut learning in deep neural net- works, Nature Machine Intelligence 2 (11) (2020) 665–673. doi:10. 1038/s42256-020-00257-z

    2020

  28. [36]

    Y. Luo, K. Guan, J. Peng, STAIR: A generic and fully-automated method to fuse multiple sources of optical satellite data to generate a high-resolution, daily and cloud-/gap-free surface reflectance prod- uct, Remote Sensing of Environment 214 (2018-09-01) 87–99. doi: 10.1016/j...

    2018 doi

  29. [37]

    Cresson, D

    R. Cresson, D. Ienco, R. Gaetano, K. Ose, D. H. Tong Minh, Optical image gap filling using deep convolutional autoencoder from optical and radar images, in: IGARSS 2019 - 2019 IEEE International Geoscience 43 and Remote Sensing Symposium, 2019-07, pp. 218–221. doi:10.1109/ IGA...

    2019

  30. [38]

    Passah, S

    A. Passah, S. N. Sur, B. Paul, D. Kandar, SAR Image Classification: A Comprehensive Study and Analysis, IEEE Access 10 (2022) 20385– 20399. doi:10.1109/ACCESS.2022.3151089

    2022

  31. [39]

    Huang, C

    Z. Huang, C. O. Dumitru, Z. Pan, B. Lei, M. Datcu, Classification of Large-Scale High-Resolution SAR Images With Deep Transfer Learning, IEEE Geoscience and Remote Sensing Letters 18 (1) (2021-01) 107–111. doi:10.1109/LGRS.2020.2965558

    2021

  32. [40]

    Gonzalez-Calabuig, M.- ´A

    M. Gonzalez-Calabuig, M.- ´A. Fern´ andez-Torres, G. Camps-Valls, Gen- erative networks for spatio-temporal gap filling of sentinel-2 reflectances, ISPRS Journal of Photogrammetry and Remote Sensing 220 (2025-02-

    2025

  33. [41]

    Y. Yuan, L. Lin, Q. Liu, R. Hang, Z.-G. Zhou, SITS-Former: A pre-trained spatio-spectral-temporal representation model for Sentinel-2 time series classification, International Journal of Applied Earth Obser- vation and Geoinformation 106 (2022-02) 102651. doi:10.1016/j.jag. 20...

    2022

  34. [42]

    doi:10.1016/j.isprsjprs.2025.01.016

    637–648. doi:10.1016/j.isprsjprs.2025.01.016

    2025 doi

  35. [43]

    Y. Yuan, L. Lin, Q. Xin, Z.-G. Zhou, Q. Liu, An Empirical Study on Data Augmentation for Pixelwise Satellite Image Time-Series Classifi- cation and Cross-Year Adaptation, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 18 (2025) 5172–5188. doi:...

    2025

  36. [44]

    Stucker, V

    C. Stucker, V. S. F. Garnot, K. Schindler, U-TILISE: A Sequence-to- Sequence Model for Cloud Removal in Optical Satellite Time Series, IEEE Transactions on Geoscience and Remote Sensing 61 (2023) 1–16. doi:10.1109/TGRS.2023.3333391

    2023

  37. [45]

    Iglesias, E

    G. Iglesias, E. Talavera, ´A. Gonz´ alez-Prieto, A. Mozo, S. G´ omez- Canaval, Data augmentation techniques in time series domain: a sur- vey and taxonomy, Neural Computing and Applications 35 (14) (2023) 10123–10145. doi:10.1007/s00521-023-08459-3

    2023 doi

  38. [46]

    A. O. de Albuquerque, O. L. F. de Carvalho, C. R. e Silva, A. S. Luiz, P. P. de Bem, R. A. T. Gomes, R. F. Guimar˜ aes, O. A. de Car- valho J´ unior, Dealing with clouds and seasonal changes for center pivot irrigation systems detection using instance segmentation in sentinel-...

    2021

  39. [47]

    Amani, A

    M. Amani, A. Ghorbanian, S. A. Ahmadi, M. Kakooei, A. Moghimi, S. M. Mirmazloumi, S. H. A. Moghaddam, S. Mahdavi, M. Ghahreman- loo, S. Parsian, et al., Google earth engine cloud computing platform for remote sensing big data applications: A comprehensive review, IEEE Journal ...

    2020

  40. [48]

    Sykas, M

    D. Sykas, M. Sdraka, D. Zografakis, I. Papoutsis, A Sentinel-2 Mul- tiyear, Multicountry Benchmark Dataset for Crop Classification and Segmentation With Deep Learning, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 15 (2022) 3323–3339. doi:10....

    2022

  41. [49]

    Voelsen, M

    M. Voelsen, M. Teimouri, F. Rottensteiner, C. Heipke, Investigating 2d and 3d convolutions for multitemporal land cover classification using remote sensing images, ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences V-3-2022 (2022) 271–279.doi: ...

    2022 doi

  42. [50]

    V. S. F. Garnot, L. Landrieu, Panoptic segmentation of satellite image time series with convolutional temporal attention networks, in: Proceed- ings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 4872–4881. doi:https://doi.org/10.48550/arXiv.2107. 07933

    2021 doi

  43. [51]

    Tarasiou, E

    M. Tarasiou, E. Chavez, S. Zafeiriou, ViTs for SITS: Vision Transform- ers for Satellite Image Time Series, in: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023-06, pp. 10418–10428. doi:10.1109/CVPR52729.2023.01004. 45

    2023

  44. [52]

    M. Majd, R. Safabakhsh, Correlational convolutional lstm for human action recognition, Neurocomputing 396 (2020) 224–229. doi:https: //doi.org/10.1016/j.neucom.2018.10.095

    2020 doi

  45. [53]

    M. Siam, S. Valipour, M. Jagersand, N. Ray, Convolutional gated re- current networks for video segmentation, in: 2017 IEEE international conference on image processing (ICIP), IEEE, 2017, pp. 3090–3094

    2017

  46. [54]

    Mobahi, M

    H. Mobahi, M. Farajtabar, P. Bartlett, Self-Distillation Amplifies Reg- ularization in Hilbert Space, in: Advances in Neural Information Pro- cessing Systems, Vol. 33, 2020, pp. 3351–3361

    2020

  47. [55]

    J. A. Chamorro Martinez, L. E. Cu´ e La Rosa, R. Q. Feitosa, I. D. Sanches, P. N. Happ, Fully convolutional recurrent networks for multi- year crop recognition from multitemporal image sequences, ISPRS Jour- nal of Photogrammetry and Remote Sensing 171 (2021-01-01) 188–201. do...

    2021 doi

  48. [56]

    M Rustowicz, R

    R. M Rustowicz, R. Cheong, L. Wang, S. Ermon, M. Burke, D. Lobell, Semantic Segmentation of Crop Type in Africa: A Novel Dataset and Analysis of Deep Learning Methods, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019, pp. 75–82

    2019

  49. [57]

    Sherstinsky, Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) network, Physica D: Nonlinear Phe- nomena 404 (2020-03-01) 132306

    A. Sherstinsky, Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) network, Physica D: Nonlinear Phe- nomena 404 (2020-03-01) 132306. doi:10.1016/j.physd.2019.132306

    2020

  50. [58]

    D. Tran, L. Bourdev, R. Fergus, L. Torresani, M. Paluri, Learning Spa- tiotemporal Features With 3D Convolutional Networks, in: Proceedings of the IEEE International Conference on Computer Vision, 2015, pp. 4489–4497. doi:10.48550/arXiv.1412.0767

    2015 doi arXiv

  51. [59]

    W. Du, D. Cˆ ot´ e, Y. Liu, SAITS: Self-attention-based imputation for time series, Expert Systems with Applications 219 (2023-06-01) 119619. doi:10.1016/j.eswa.2023.119619. 46

    2023

  52. [60]

    Rehmer, A

    A. Rehmer, A. Kroll, On the vanishing and exploding gradient prob- lem in Gated Recurrent Units, IF AC-PapersOnLine 53 (2) (2020-01-01) 1243–1248. doi:10.1016/j.ifacol.2020.12.1342

    2020 doi

  53. [796]

    doi:10.3390/w9100796

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.