Pith. sign in

REVIEW 5 major objections 5 minor 51 references

Spatial-Temporal-Spectral Mamba with Sparse Deformable Token Sequence for Enhanced MODIS Time Series Classification

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A sparse deformable token sequence lets a Mamba model classify MODIS time series more accurately at lower cost.

desk verdict New Mamba-based architecture for MODIS time series land cover mapping, but headline accuracies rest on an undefined 'filtered ground truth' and need verification. read the letter →

arxiv 2508.02839 v1 pith:JGG5THAV submitted 2025-07-29 eess.IV eess.SP

Zack Dewis , Zhengsen Xu , Yimin Zhu , Motasem Alkayid , Mabel Heffring , Lincoln Linlin Xu This is my paper
classification eess.IVeess.SP
keywords spatial-temporal-spectralMambaMODIStimeseriesclassificationlandcoversparsedeformabletokensequencestatespacemodelremotesensingsequencing
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 tries to establish that a Mamba-style state-space sequence model can classify coarse-resolution MODIS time series more accurately and more cheaply than prior CNN, Transformer, RNN, and Mamba baselines, provided the token sequence is made sparse, deformable, and split by spatial, spectral, and temporal dimensions. The proposed STSMamba first disentangles temporal and spectral information with a temporal grouped stem, then learns which tokens matter and feeds only a short reordered subset through three dedicated Mamba modules. On Saskatchewan MODIS data it reports 97.59% overall accuracy, and when trained on Saskatchewan and transferred to Alberta it reports 82.36% overall accuracy, both above all listed baselines. If true, this points to a way of using Mamba sequence models on long, coupled satellite time series without paying the redundancy and correlation-decay costs of long dense token sequences.

What carries the argument

The load-bearing mechanism is a sparse, deformable Mamba sequencing (SDMS) scheme. Rather than feeding all tokens in a fixed scan order, the model computes an attention matrix over one dimension, sparsifies it to keep only the most relevant tokens, and reorders the survivors into a short Mamba input; the same idea is applied in three modules—sparse deformable temporal Mamba (SDTM), sparse deformable spectral Mamba (SDSpeM), and sparse deformable spatial Mamba (SDSpaM)—each explicitly targeting one source of information in MODIS. A temporal grouped stem (TGS) feeds the modules by processing each time step separately, which prevents spectral and temporal channels from being mixed before feature learning. The stated effect is linear-complexity long-range modeling without redundant tokens or long-range correlation decay.

What would settle it

Run the same train/test protocol on the unfiltered Saskatchewan and Alberta ground truth with the same baselines; if STSMamba's overall-accuracy lead over MambaHSI and SwinT shrinks below the reported margins, or if its computational cost at equal token counts is not below dense Mamba, the claim that sparse deformable sequencing is what improves accuracy and efficiency is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the bottleneck in MODIS time-series classification is not model capacity but the Mamba token sequence: long, mixed, dense, and rigid sequences entangle spatial, temporal, and spectral information and carry redundant tokens, which wastes computation and dilutes class signatures. STSMamba is the architecture built to test that claim. It separates the data cube by time step, uses a temporal grouped stem to learn 18 spectral-spatial features per time step, and then constructs three short Mamba sequences—temporal, spectral, and spatial—each reduced by a learnable sparse attention mask to the handful of most relevant tokens (six temporal tokens in the reported configuration). The paper reports that this design outperforms RNN, CNN, Transformer, and existing Mamba baselines on the Saskatchewan dataset (97.59% overall accuracy, 96.76% Kappa) and transfers better to Alberta (82.36% overall accuracy), with sparsity ablations showing temporal sparsity 0.3 and spectral sparsity 0.8 as the best operating points.

Load-bearing premise

The evaluation depends on the 'filtered ground truth' and the train/test split being representative, and the paper never specifies what was filtered, so selective removal of mixed or ambiguous pixels could make the reported accuracies easier to achieve.

Editorial extensions

If this is right

  • On the Saskatchewan benchmark, STSMamba's 97.59% overall accuracy is 4.55 points above MambaHSI and roughly 3.8 points above SwinT, so even recent strong baselines leave room on coarse-resolution time series.
  • Trained on Saskatchewan and tested on Alberta, STSMamba's 82.36% overall accuracy beats all listed baselines, implying the learned class signatures transfer across eco-regions better than prior models.
  • The sparsity ablation indicates that temporal compression to 0.3 helps while spectral compression beyond 0.8 hurts, so redundancy is concentrated in time rather than in spectral bands.
  • Because only six temporal tokens are fed into the temporal Mamba module, the claimed computational savings scale with the sparsity ratio, making the model more feasible for large-area MODIS mapping.
  • The three dedicated sparse deformable modules jointly support the argument that explicit spatial, spectral, and temporal modeling is what recovers subtle class signatures in mixed pixels.

Reading between the lines

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

  • Beyond the paper, the decouple-then-sparsify recipe could generalize to other dense multi-temporal sensors, since any long satellite time series faces the same redundancy and correlation-decay problem.
  • An open question the paper leaves implicit is whether the sparse token selections are stable across regions and seasons; if the selected tokens shift sharply under domain shift, part of the Alberta generalization gain may come from the deformability itself rather than from sparsity.
  • A natural testable extension would compare STSMamba against a dense Mamba variant with the same token count, isolating whether the accuracy gain comes from selecting the right tokens or simply from using fewer of them.
  • The undefined 'filtered ground truth' means the absolute accuracies may include only easier pixels; the relative ordering of methods is the more robust claim to carry forward.
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

5 major / 5 minor

Summary. This paper presents STSMamba, a Mamba-based classifier for MODIS time series that decomposes the input into spatial, spectral, and temporal Mamba branches with a sparse, deformable token sequencing mechanism and a temporal grouped stem. On the Saskatchewan MODIS dataset it reports OA 97.59%, AA 96.01%, and Kappa 96.76%, outperforming ten baselines, and on the Alberta dataset it reports OA 82.36%, again above all listed baselines. The paper claims reduced computational complexity and includes an ablation over sparsity ratios. The central evidence is empirical; no code or data is provided.

Significance. The proposed architecture is well motivated by the MODIS-specific challenges of mixed pixels and spatio-temporal-spectral coupling, and the cross-region generalization experiment from Saskatchewan to Alberta is a valuable evaluation design. The sparse deformable Mamba sequencing idea is interesting and, if validated, would be a useful contribution to time-series remote sensing classification. However, the headline accuracy claim currently rests on an undefined 'filtered ground truth' and on single-run evaluations with no error bars, so the significance cannot be assessed until the evaluation is clarified and strengthened.

major comments (5)
  1. [Tables II and IV] The table captions state that results are computed on 'filtered ground truth', but no definition of the filtering operation appears in Section IV. It is not stated which pixels are excluded, by what criterion, in what proportion, or whether the same filter is applied to training and test partitions and to all baselines. Because the Introduction identifies mixed pixels as a core MODIS difficulty, a filter that removes mixed or transition-zone pixels would preferentially remove the hard cases the method is intended to handle and could inflate the reported OA/AA/Kappa and alter the method ranking. Please define the filter, report unfiltered results for all methods, and describe the filtering effect per class.
  2. [Section IV-C/D; Tables II and IV] All accuracy numbers are reported as single runs with no error bars, confidence intervals, or significance tests. On Saskatchewan the margin over the best baseline (SSRN, 96.28%) is only 1.31 OA points, and on Alberta the margin over GRU (80.45%) is 1.91 OA points; these margins are comparable to typical run-to-run variation in deep classifiers. The paper also omits per-class sample sizes and the train/test split proportions, without which the class-wise entries in Table IV (e.g., SwinT with 7.50% on Taiga-needleleaf) cannot be interpreted. Please provide repeated-seed results and a statistical comparison.
  3. [Section IV-E; Table III] The sparsity ratios appear to be selected by comparing OA on the evaluation data: the text concludes that temporal sparse ratio 0.3 and spectral sparse ratio 0.8 are optimal, and this configuration is then used in Tables II and IV. No validation split is mentioned, and the OA differences between ratios in Table III are as small as 0.04–0.17 percentage points. If the ratios were selected on the same test set, the reported headline numbers are the result of test-set selection rather than an independent evaluation. Please use a validation split, report the selection procedure, and give all configurations tried.
  4. [Abstract and Section V] The abstract and conclusion claim 'reduced computational complexity', but the manuscript reports no parameter counts, FLOPs, inference time, memory usage, or training time for STSMamba or any baseline. Shortening the temporal Mamba sequence from T=23 to six tokens in SDTM does not by itself establish overall complexity, because the sparse token selection itself requires an attention matrix computation and a selection operation. Please add quantitative complexity measurements for the full model and the baselines.
  5. [Abstract and Section V] The paper claims 'extensive ablation studies' to justify the key building blocks, but the only ablation table is Table III, which varies the sparsity ratios. There is no ablation removing or replacing the temporal grouped stem (TGS) or any of the three modules SDSpaM, SDSpeM, and SDTM. Without these ablations, the contribution of each architectural component to the final accuracy is not demonstrated. Please add component-wise ablations (or explicitly state which components were ablated and report those results).
minor comments (5)
  1. [Section IV-B] The baseline training protocol is not described: there are no optimizer settings, learning rates, schedules, epochs, or data augmentation choices for the ten baselines. A statement on whether the same training budget and hyperparameter search were used for all methods would make the comparison credible.
  2. [Figure 3] The caption of Figure 3 contains 'train Saskatchewan dataset'; the grammar should be corrected, and the axis labels and units should be legible.
  3. [Equation (2)] The mathematical notation in the token-selection equations is typeset with missing symbols in the available text; please ensure the camera-ready PDF renders all matrices, softmax, and sparse-selection operations clearly, and define the 'six re-ordered tokens' selection algorithm explicitly.
  4. [General] There is no data or code availability statement. For an empirical benchmark paper, releasing the data-processing and filter code and the training scripts would substantially improve reproducibility.
  5. [Section II] Minor typo: 'sequance' should be 'sequence' in the related-work discussion of long Mamba sequences.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the paper is an empirical architecture benchmark whose accuracy claims are measured outputs, not derived quantities; the main weakness is an undefined 'filtered ground truth' in Tables II and IV, which is an evaluation-transparency issue rather than a circular reduction.

full rationale

This is an empirical benchmark paper, not a derivation: the central claim is measured classification accuracy on MODIS time series data, and the reported OA/AA/Kappa values in Tables II and IV are network outputs evaluated on test pixels, not quantities algebraically derived from fitted inputs. The methodology (TGS, SDMS, SDSpaM/SDSpeM/SDTM) builds on standard Mamba, convolution, and attention operations defined from the input tensor; no equation in Section III defines a predicted accuracy in terms of a fitted parameter. The sparsity-ratio sensitivity study in Table III is a hyperparameter search: if conducted on the same evaluation set it can inflate results through selection, but it is not a 'fitted input called prediction' because the reported accuracies are still measured on data rather than forced by construction. The 'filtered ground truth' wording in the titles of Tables II and IV is the paper's most serious weakness, but it is a missing-support/selection-bias concern: the filtering rule is never defined in Section IV, so the reported numbers may not be reproducible and may reflect an easier subset, yet the filter is not an input to any derivation that produces the accuracy values. The only possible self-citation is reference [27], which lists 'L. Xu' as a co-author and is used only as RNN/LSTM background; it is not load-bearing. No uniqueness theorem or ansatz is imported from the authors' prior work, and no prediction reduces by construction to its own input. Therefore the circularity score is low, reflecting only minor transparency and non-load-bearing self-citation concerns, not actual circularity.

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

The empirical claims rest on hand-chosen sparsity ratios and several domain assumptions about Mamba behavior and MODIS ground truth. No new physical entities are introduced; the sparse deformable sequence is a mechanism, not an entity.

free parameters (3)
  • Temporal sparsity ratio = 0.3
    Chosen from ablation results as optimal in Section IV-E; it controls that 6 of the 23 temporal tokens are retained in the SDTM module.
  • Spectral sparsity ratio = 0.8
    Chosen from ablation results as optimal in Section IV-E; it controls the number of retained spectral tokens in the SDSpeM module.
  • Stem output channels = 18
    Hand-set number of feature channels after the temporal grouped stem in Section III-B; no sensitivity analysis is provided for this choice.
assumptions (4)
  • domain assumption Mamba state-space blocks capture long-range dependencies with linear computational complexity.
    Assumed from prior Mamba literature [34]-[37]; not re-derived in this paper.
  • domain assumption Attention-based sparsification selects the most class-relevant tokens.
    Central to the SDMS design in Section III-C; if the six selected tokens omit key phenological cues, the accuracy claims collapse.
  • domain assumption The filtered ground truth is accurate and representative.
    All metrics depend on this ground truth, but the filtering rule is not defined in Section IV.
  • domain assumption Spectral and temporal information can be safely decoupled early in the network.
    The temporal grouped stem in Section III-B separates time steps before spectral mixing, assuming the coupling is harmful rather than useful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatial-Temporal-Spectral Mamba with Sparse Deformable Token Sequence for Enhanced MODIS Time Series Classification." pith.science (2026). https://pith.science/paper/JGG5THAV

@misc{pith2026250802839,
  author       = {Pith},
  title        = {Pith review of: Spatial-Temporal-Spectral Mamba with Sparse Deformable Token Sequence for Enhanced MODIS Time Series Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGG5THAV}},
  note         = {Machine review of arXiv:2508.02839}
}
read the original abstract

Although MODIS time series data are critical for supporting dynamic, large-scale land cover land use classification, it is a challenging task to capture the subtle class signature information due to key MODIS difficulties, e.g., high temporal dimensionality, mixed pixels, and spatial-temporal-spectral coupling effect. This paper presents a novel spatial-temporal-spectral Mamba (STSMamba) with deformable token sequence for enhanced MODIS time series classification, with the following key contributions. First, to disentangle temporal-spectral feature coupling, a temporal grouped stem (TGS) module is designed for initial feature learning. Second, to improve Mamba modeling efficiency and accuracy, a sparse, deformable Mamba sequencing (SDMS) approach is designed, which can reduce the potential information redundancy in Mamba sequence and improve the adaptability and learnability of the Mamba sequencing. Third, based on SDMS, to improve feature learning, a novel spatial-temporal-spectral Mamba architecture is designed, leading to three modules, i.e., a sparse deformable spatial Mamba module (SDSpaM), a sparse deformable spectral Mamba module (SDSpeM), and a sparse deformable temporal Mamba module (SDTM) to explicitly learn key information sources in MODIS. The proposed approach is tested on MODIS time series data in comparison with many state-of-the-art approaches, and the results demonstrate that the proposed approach can achieve higher classification accuracy with reduced computational complexity.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 43 canonical work pages

  1. [1]

    Modis collection 5 global land cover: Algorithm refinements and characterization of new datasets,

    M. A. Friedl, D. Sulla-Menashe, B. Tan, A. Schneider, N. Ramankutty, A. Sibley, and X. Huang, “Modis collection 5 global land cover: Algorithm refinements and characterization of new datasets,” Remote Sensing of Environment , vol. 114, no. 1, pp. 168–182, 2010. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0034425709002673

    work page 2010

  2. [2]

    Cropland distributions from temporal unmixing of modis data,

    D. B. Lobell and G. P. Asner, “Cropland distributions from temporal unmixing of modis data,” Remote sensing of Environment, vol. 93, no. 3, pp. 412–422, 2004

    work page 2004

  3. [3]

    The pixel: a snare and a delusion,

    P. Fisher, “The pixel: a snare and a delusion,” International Journal of remote sensing, vol. 18, no. 3, pp. 679–685, 1997

    work page 1997

  4. [4]

    Effect of spatial resolution on classification errors of pure and mixed pixels in remote sensing,

    P.-F. Hsieh, L. Lee, and N.-Y . Chen, “Effect of spatial resolution on classification errors of pure and mixed pixels in remote sensing,” IEEE Transactions on Geoscience and Remote Sensing , vol. 39, no. 12, pp. 2657–2663, 2001

    work page 2001

  5. [5]

    The use of single-date modis imagery for estimating large-scale urban impervious surface fraction with spectral mixture analysis and machine learning techniques,

    C. Deng and C. Wu, “The use of single-date modis imagery for estimating large-scale urban impervious surface fraction with spectral mixture analysis and machine learning techniques,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 86, pp. 100–110, 2013. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0924271613002116

    work page 2013

  6. [6]

    A unified spatial-spectral-temporal fusion model using landsat and modis imagery,

    B. Chen and B. Xu, “A unified spatial-spectral-temporal fusion model using landsat and modis imagery,” in 2014 Third International Workshop on Earth Observation and Remote Sensing Applications (EORSA), 2014, pp. 256–260

    work page 2014

  7. [7]

    Comparison of support vector machine, neural network, and cart algorithms for the land-cover classification using limited training data points,

    Y . Shao and R. S. Lunetta, “Comparison of support vector machine, neural network, and cart algorithms for the land-cover classification using limited training data points,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 70, pp. 78–87, 2012

    work page 2012

  8. [8]

    Land cover classification with support vector machine applied to modis imagery,

    P. Gonc ¸alves, H. Carr ˜ao, A. Pinheiro, M. Caetano et al. , “Land cover classification with support vector machine applied to modis imagery,” Global developments in environmental earth observation from space, pp. 517–525, 2006

    work page 2006

Show all 51 references
  1. [9]

    Vuolo and C

    F. Vuolo and C. Atzberger, “Exploiting the classification performance of support vector machines with multi-temporal moderate-resolution imaging spectroradiometer (modis) data in areas of agreement and JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 9 TABLE IV SPA...

    2014

  2. [10]

    Charac- terizing land cover/land use from multiple years of landsat and modis time series: A novel approach using land surface phenology modeling and random forest classifier,

    L. H. Nguyen, D. R. Joshi, D. E. Clay, and G. M. Henebry, “Charac- terizing land cover/land use from multiple years of landsat and modis time series: A novel approach using land surface phenology modeling and random forest classifier,” Remote sensing of environment , vol. 238,...

    2020

  3. [11]

    Temporal optimisation of image acquisition for land cover classification with random forest and modis time-series,

    I. Nitze, B. Barrett, and F. Cawkwell, “Temporal optimisation of image acquisition for land cover classification with random forest and modis time-series,” International Journal of Applied Earth Observation and Geoinformation, vol. 34, pp. 136–146, 2015

    2015

  4. [12]

    Developing a random forest algorithm for modis global burned area classification,

    R. Ramo and E. Chuvieco, “Developing a random forest algorithm for modis global burned area classification,” Remote Sensing, vol. 9, no. 11, p. 1193, 2017

    2017

  5. [13]

    Mapping annual land use and land cover changes using modis time series,

    H. Yin, D. Pflugmacher, R. E. Kennedy, D. Sulla-Menashe, and P. Hostert, “Mapping annual land use and land cover changes using modis time series,” IEEE Journal of selected topics in applied earth observations and remote sensing , vol. 7, no. 8, pp. 3421–3427, 2014

    2014

  6. [14]

    Feature selection of time series modis data for early crop classification using random forest: A case study in kansas, usa,

    P. Hao, Y . Zhan, L. Wang, Z. Niu, and M. Shakir, “Feature selection of time series modis data for early crop classification using random forest: A case study in kansas, usa,” Remote Sensing, vol. 7, no. 5, pp. 5347–5369, 2015

    2015

  7. [15]

    A cascaded spectral–spatial cnn model for super-resolution river mapping with modis imagery,

    Z. Yin, F. Ling, X. Li, X. Cai, H. Chi, X. Li, L. Wang, Y . Zhang, and Y . Du, “A cascaded spectral–spatial cnn model for super-resolution river mapping with modis imagery,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–13, 2021

    2021

  8. [16]

    Classification for remote sensing data with improved cnn-svm method,

    X. Sun, L. Liu, C. Li, J. Yin, J. Zhao, and W. Si, “Classification for remote sensing data with improved cnn-svm method,” Ieee Access, vol. 7, pp. 164 507–164 516, 2019

    2019

  9. [17]

    Spatiotemporal satellite image fusion using deep convolutional neural networks,

    H. Song, Q. Liu, G. Wang, R. Hang, and B. Huang, “Spatiotemporal satellite image fusion using deep convolutional neural networks,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 11, no. 3, pp. 821–829, 2018

    2018

  10. [18]

    Large area crops mapping by phenological horizon attention transformer (phat) method using modis time-series imagery,

    Q. Gao, T. Wu, H. Tang, J. Yang, and S. Wang, “Large area crops mapping by phenological horizon attention transformer (phat) method using modis time-series imagery,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 2025

    2025

  11. [19]

    High-resolution seamless mapping of the leaf area index via multisource data and the transformer deep learning model,

    P. Chen, K. Zhou, and H. Fang, “High-resolution seamless mapping of the leaf area index via multisource data and the transformer deep learning model,” IEEE Transactions on Geoscience and Remote Sensing, 2025

    2025

  12. [20]

    Msnet: A multi-stream fusion network for remote sensing spatiotemporal fusion based on transformer and convolution,

    W. Li, D. Cao, Y . Peng, and C. Yang, “Msnet: A multi-stream fusion network for remote sensing spatiotemporal fusion based on transformer and convolution,” Remote Sensing, vol. 13, no. 18, p. 3724, 2021

    2021

  13. [21]

    Spectral–spatial–temporal transformers for hyperspectral image change detection,

    Y . Wang, D. Hong, J. Sha, L. Gao, L. Liu, Y . Zhang, and X. Rong, “Spectral–spatial–temporal transformers for hyperspectral image change detection,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–14, 2022

    2022

  14. [22]

    A cnn-lstm model for soil organic carbon content prediction with long time series of modis-based phenological variables,

    L. Zhang, Y . Cai, H. Huang, A. Li, L. Yang, and C. Zhou, “A cnn-lstm model for soil organic carbon content prediction with long time series of modis-based phenological variables,” Remote sensing, vol. 14, no. 18, p. 4441, 2022

    2022

  15. [23]

    Application of long short-term memory neural network model for the reconstruction of modis land surface tem- perature images,

    N. Arslan and A. Sekertekin, “Application of long short-term memory neural network model for the reconstruction of modis land surface tem- perature images,” Journal of Atmospheric and Solar-Terrestrial Physics, vol. 194, p. 105100, 2019

    2019

  16. [24]

    Development of the glass 250-m leaf area index product (version 6) from modis data using the bidirectional lstm deep learning model,

    H. Ma and S. Liang, “Development of the glass 250-m leaf area index product (version 6) from modis data using the bidirectional lstm deep learning model,” Remote sensing of environment , vol. 273, p. 112985, 2022

    2022

  17. [25]

    Learning spectral-spatial- temporal features via a recurrent convolutional neural network for change detection in multispectral imagery,

    L. Mou, L. Bruzzone, and X. X. Zhu, “Learning spectral-spatial- temporal features via a recurrent convolutional neural network for change detection in multispectral imagery,” IEEE Transactions on Geo- science and Remote Sensing , vol. 57, no. 2, pp. 924–935, 2018

    2018

  18. [26]

    Land cover classification via multitemporal spatial data by deep recurrent neural networks,

    D. Ienco, R. Gaetano, C. Dupaquier, and P. Maurel, “Land cover classification via multitemporal spatial data by deep recurrent neural networks,” IEEE Geoscience and Remote Sensing Letters , vol. 14, no. 10, pp. 1685–1689, 2017

    2017

  19. [27]

    Land cover classification using lstm on time series landsat data,

    Y . Sun, B. Chen, and L. Xu, “Land cover classification using lstm on time series landsat data,” Remote Sensing, vol. 11, no. 21, p. 2560, 2019

    2019

  20. [28]

    A semi-supervised deep learning approach for time series land cover classification,

    W. Jing and L. Chao, “A semi-supervised deep learning approach for time series land cover classification,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 160, pp. 192–206, 2020

    2020

  21. [29]

    Temporal convolutional neural network for the classification of satellite image time series,

    C. Pelletier, G. I. Webb, and F. Petitjean, “Temporal convolutional neural network for the classification of satellite image time series,” Remote Sensing, vol. 11, no. 5, p. 523, 2019

    2019

  22. [30]

    A review of temporal convolutional net- works for land use classification from satellite time series,

    A. L. Brock and A. Abdallah, “A review of temporal convolutional net- works for land use classification from satellite time series,” Computers and Electronics in Agriculture , vol. 199, p. 107075, 2022

    2022

  23. [31]

    The use of remote sensing and gis in watershed level analyses of non-point source pollution problems,

    P. Basnyat, L. D. Teeter, B. G. Lockaby, and K. M. Flynn, “The use of remote sensing and gis in watershed level analyses of non-point source pollution problems,” Forest Ecology and Management, vol. 128, no. 1-2, pp. 65–73, 2000

    2000

  24. [32]

    Transformer-based land use and land cover classification with explain- ability using satellite imagery,

    M. Khan, A. Hanan, M. Kenzhebay, M. Gazzea, and R. Arghandeh, “Transformer-based land use and land cover classification with explain- ability using satellite imagery,” Scientific Reports , vol. 14, no. 1, p. 16744, 2024

    2024

  25. [33]

    Earthformer: Exploring space-time transformers for earth system fore- casting,

    Z. Gao, X. Shi, H. Wang, Y . Zhu, Y . B. Wang, M. Li, and D.-Y . Yeung, “Earthformer: Exploring space-time transformers for earth system fore- casting,” Advances in Neural Information Processing Systems , vol. 35, pp. 25 390–25 403, 2022

    2022

  26. [34]

    Mambahsi: Spa- tial–spectral mamba for hyperspectral image classification,

    Y . Li, Y . Luo, L. Zhang, Z. Wang, and B. Du, “Mambahsi: Spa- tial–spectral mamba for hyperspectral image classification,” IEEE Trans- actions on Geoscience and Remote Sensing , vol. 62, pp. 1–16, 2024

    2024

  27. [35]

    3dss-mamba: 3d-spectral- spatial mamba for hyperspectral image classification,

    Y . He, B. Tu, B. Liu, J. Li, and A. Plaza, “3dss-mamba: 3d-spectral- spatial mamba for hyperspectral image classification,” IEEE Transac- tions on Geoscience and Remote Sensing , vol. 62, pp. 1–16, 2024

    2024

  28. [36]

    Hypermamba: A spectral- spatial adaptive mamba for hyperspectral image classification,

    Q. Liu, J. Yue, Y . Fang, S. Xia, and L. Fang, “Hypermamba: A spectral- spatial adaptive mamba for hyperspectral image classification,” IEEE Transactions on Geoscience and Remote Sensing , 2024

    2024

  29. [37]

    A comprehensive survey for hyperspectral image classification: The evolution from conventional to transformers and mamba models,

    M. Ahmad, S. Distifano, A. M. Khan, M. Mazzara, C. Li, H. Li, J. Aryal, Y . Ding, G. Vivone, and D. Hong, “A comprehensive survey for hyperspectral image classification: The evolution from conventional to transformers and mamba models,” arXiv preprint arXiv:2404.14955 , 2024. ...

    2024 arXiv

  30. [38]

    The new generation brain- inspired sparse learning: A comprehensive survey,

    I. K. Sinha, S. Verma, and K. P. Singh, “The new generation brain- inspired sparse learning: A comprehensive survey,” IEEE Transactions on Artificial Intelligence , vol. PP, no. 99, pp. 1–1, 2022

    2022

  31. [39]

    Sparsity in transformers: A systematic literature review,

    M. Farina, U. Ahmad, A. Taha, H. Younes, Y . Mesbah, X. Yu, and W. Pedrycz, “Sparsity in transformers: A systematic literature review,” Neurocomputing, p. 127468, 2024

    2024

  32. [40]

    Generating long sequences with sparse transformers,

    R. Child, S. Gray, A. Radford, and I. Sutskever, “Generating long sequences with sparse transformers,” arXiv preprint arXiv:1904.10509 , 2019

    1904 arXiv

  33. [41]

    Efficient content- based sparse attention with routing transformers,

    A. Roy, M. Saffar, A. Vaswani, and D. Grangier, “Efficient content- based sparse attention with routing transformers,” Transactions of the Association for Computational Linguistics , vol. 9, pp. 53–68, 2021

    2021

  34. [42]

    Sparse is enough in scaling trans- formers,

    S. Jaszczur, A. Chowdhery, A. Mohiuddin, L. Kaiser, W. Gajewski, H. Michalewski, and J. Kanerva, “Sparse is enough in scaling trans- formers,” Advances in Neural Information Processing Systems , vol. 34, pp. 9895–9907, 2021

    2021

  35. [43]

    Adaptively sparse transformers,

    G. M. Correia, V . Niculae, and A. F. Martins, “Adaptively sparse transformers,” arXiv preprint arXiv:1909.00015 , 2019

    1909 arXiv

  36. [44]

    Sparseswin: Swin transformer with sparse transformer block,

    K. Pinasthika, B. S. P. Laksono, R. B. P. Irsal, S. Shabiyya, and N. Yudistira, “Sparseswin: Swin transformer with sparse transformer block,” Neurocomputing, vol. 580, p. 127433, 2024

    2024

  37. [45]

    Exphormer: Sparse transformers for graphs,

    H. Shirzad, A. Velingker, B. Venkatachalam, D. J. Sutherland, and A. K. Sinop, “Exphormer: Sparse transformers for graphs,” in International Conference on Machine Learning . PMLR, 2023, pp. 31 613–31 632

    2023

  38. [46]

    Deformable convnets v2: More deformable, better results,

    X. Zhu, H. Hu, S. Lin, and J. Dai, “Deformable convnets v2: More deformable, better results,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 9308–9316

    2019

  39. [47]

    Internimage: Exploring large-scale vision foundation models with deformable convolutions,

    W. Wang, J. Dai, Z. Chen, Z. Huang, Z. Li, X. Zhu, X. Hu, T. Lu, L. Lu, H. Li et al., “Internimage: Exploring large-scale vision foundation models with deformable convolutions,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 14...

    2023

  40. [48]

    Deformable detr: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to-end object detection,”arXiv preprint arXiv:2010.04159, 2020

    2010 arXiv

  41. [49]

    Vision transformer with deformable attention,

    Z. Xia, X. Pan, S. Song, L. E. Li, and G. Huang, “Vision transformer with deformable attention,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 4794–4803

    2022

  42. [50]

    Dunet: A deformable network for retinal vessel segmentation,

    Q. Jin, Z. Meng, T. D. Pham, Q. Chen, L. Wei, and R. Su, “Dunet: A deformable network for retinal vessel segmentation,” Knowledge-Based Systems, vol. 178, pp. 149–162, 2019

    2019

  43. [51]

    Circa 2010 land cover of canada: Local optimization methodology and product development,

    R. Latifovic, D. Pouliot, and I. Olthof, “Circa 2010 land cover of canada: Local optimization methodology and product development,” Remote Sensing, vol. 9, p. 1098, 10 2017

    2010

Pith tools

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