Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

$T^{3}S$: Think in Thermal Time for Generalizable Crop Mapping from Satellite Image Time Series

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper claims that replacing calendar-date sampling of satellite image time series with sampling by cumulative growing degree days (thermal time) makes crop-type classifiers generalize across growing seasons.

desk verdict A simple, well-ablated thermal-time sampling trick that gains ~6 points on cross-year crop mapping, but the single base-temperature assumption and a stale abstract need referee attention. read the letter →

arxiv 2506.12885 v4 pith:2DMEWCES submitted 2025-06-15 cs.CV eess.IV

classification cs.CVeess.IV
keywords croptypeclassificationthermaltimegrowingdegreedayssatelliteimageseriescross-yeargeneralizationuncertaintycalibrationearly-seasontemporalsampling
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

This paper claims that a purely preprocessing change—replacing calendar-date sampling of satellite image time series with sampling by cumulative growing degree days (thermal time)—makes crop-type classifiers generalize across growing seasons. On a new country-wide Swiss dataset spanning 2021–2023, T3S beats strong baselines in leave-one-year-out tests by about 6 percentage points in overall accuracy while cutting expected calibration error roughly fourfold. The gains hold under scarce labels and in early-season prediction, with no architectural change to the underlying model. The intended contribution is a simple, model-agnostic way to inject ecophysiological knowledge into any satellite time series model.

What carries the argument

The central object is the thermal-time sampling grid: for each day, growing degree days are computed as max(0, (Tmax+Tmin)/2 − Tbase) with Tbase = 0 °C, accumulated across the season to a cumulative GDD curve. The full cGDD range is divided into T uniform intervals, and within each interval the satellite observation with the fewest cloud-contaminated pixels is selected. This procedure replaces the calendar grid (e.g., one image every N days) with a grid of equal heat increments, so that early warm years and late cool years map onto the same sequence positions. The mechanism works because model training sees sequences whose positions correspond to biologically comparable phases, and because cloud-aware selection removes noisy observations without distorting the temporal grid.

What would settle it

A direct test would be to apply T3S to a crop type with a documented higher base temperature, such as maize (commonly around 10 °C), and compare cross-year accuracy against the 0 °C version; if accuracy does not improve or drops relative to calendar sampling for that crop while improving for cereals, the single-base-temperature assumption is violated. Alternatively, an analysis of class-wise accuracy in the released SwissCrop predictions would show whether the most heat-sensitive crops fail to benefit.

Watch

Extended reading notes

Core claim

The central discovery is that the temporal grid itself, not the model or the positional encoding, is the main carrier of cross-year generalization. Uniform subsampling in calendar time leaves inter-annual phenology shifts in the input, because the same calendar date corresponds to different growth stages in different years. Re-indexing the sequence so that each of the T=24 input slots covers an equal increment of cumulative growing degree days, and selecting the least cloudy image in each thermal bin, aligns phenologically equivalent stages across years. Trained on one year and tested on other years, this re-indexing yields about 77% overall accuracy versus about 71% for the U-TAE baseline, with expected calibration error falling from about 0.05 to about 0.01. The paper also finds that adding thermal time as a positional encoding gives only modest gains, suggesting the sampling grid, not the embedding, is what realigns the data.

Load-bearing premise

The method assumes that a single base temperature of 0 °C in the growing-degree-day formula adequately represents the phenological development of all 50 crop types across Switzerland; if some crops develop on a different heat threshold or are limited by water rather than temperature, thermal bins would misalign their growth stages.

Editorial extensions

If this is right

  • Because T3S is model-agnostic, any satellite-time-series crop classifier can adopt it by swapping the sampling grid; the paper reports consistent gains across several architecturally distinct backbones, so the effect is not tied to one model.
  • Operational users can produce trustworthy crop maps for the current season without current-year labels: with only 10% of training labels, T3S outperforms the full-data baseline in accuracy and calibration, and by end of June it matches the full-season baseline.
  • Calibration improvement means confidence scores become usable for downstream decisions: expected calibration error drops to roughly 0.01, approaching perfect calibration, so low-confidence regions can be flagged for ground checking.
  • Early-season mapping becomes practical: truncating test series at the 50th percentile (end of June) leaves T3S at accuracy and calibration comparable to the baseline's full-season performance.

Reading between the lines

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

  • If thermal re-indexing is the causal mechanism, the method should transfer to other temperature-driven biological mapping tasks, such as grassland cutting detection, forest phenology, or pest-pressure windows, wherever accumulated heat is the pacing variable; this is an extension the authors mention only in passing.
  • A testable extension is to replace the fixed base temperature of 0 °C with an estimated per-class or per-pixel base temperature, which would likely sharpen gains on warm-season crops while keeping the method preprocessing-only.
  • Because the comparison shows thermal positional encoding alone gives little gain, we infer that future phenology-aware models should focus on input re-indexing rather than on richer positional embeddings.
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

3 major / 5 minor

Summary. This paper proposes Thermal-Time-based Temporal Sampling (T3S), a preprocessing method that re-indexes satellite image time series by cumulative growing degree days (cGDD): each sequence is divided into T uniform intervals in cGDD and the least cloudy observation in each interval is selected. The authors evaluate T3S on SwissCrop (2021–2023, 50 crop classes) with a U-TAE backbone under leave-one-year-out protocols, reporting roughly 6% absolute accuracy gain over U-TAE, about a fourfold ECE reduction, and improved low-data and early-season behavior. The method is deterministic, requires no architectural changes, and the paper releases a new dataset and code.

Significance. T3S is a simple, preprocessing-only intervention with no fitted parameters, and the Deformable Sampling ablation cleanly separates cloud-selection effects from thermal re-timing. The six train/test year-pair evaluation shows consistent accuracy and calibration gains across folds, and the public SwissCrop dataset with paired temperature data is a valuable resource for cross-year crop mapping research. The main limitations are that the empirical evidence covers only one backbone and one country, the thermal index uses a single base temperature without sensitivity analysis, and the spatial granularity of the thermal signal is ambiguous.

major comments (3)
  1. [Abstract and Section 5 (Table 2)] The abstract claims evaluation on three architecturally distinct backbones and on the cross-region TimeMatch benchmark spanning Denmark and France, but the full text contains only U-TAE results on SwissCrop (Table 2). No other backbone or cross-region experiment appears anywhere in the main text, tables, or appendices. Because 'model-agnostic' and cross-region generalization are central claims in the title and contributions, this mismatch is load-bearing. Either add the missing experiments or revise the abstract and contribution statements to reflect the actual scope, and at minimum add one additional backbone to support the model-agnostic claim.
  2. [Section 3.3, Eq. (1)] The GDD computation uses a single base temperature Tbase = 0 °C uniformly for all 50 crop types in SwissCrop. The paper provides no sensitivity analysis over Tbase or comparison with crop-specific base temperatures, even though warm-season crops such as maize commonly require higher base temperatures. Since the mechanism is that thermal bins align phenologically equivalent growth stages, a miscalibrated thermal index could shift bins toward uninformative periods and the reported gains could shrink or reverse. Please add a sensitivity analysis over Tbase (e.g., 0, 2, 5, 8, 10 °C) or crop-group-specific bases, and report how class-wise results change.
  3. [Section 3.3, Algorithm 1 and Section 4] Algorithm 1 takes cGDD as a length-L vector (line 9 uses cGDD[i]), but Section 4 states that MeteoSwiss temperature data is resampled to 10 m to match Sentinel-2 imagery. It is therefore unclear whether thermal intervals are computed per pixel, per data cube, or from a single country-wide average temperature (Figure 1 shows 'spatially averaged cGDD across Switzerland'). If a single national curve is used, local temperature variation in a topographically diverse country is ignored and the phenological alignment claim is weakened; if per-pixel cGDD is intended, the algorithm and its notation need a spatial dimension. Please clarify this and, if feasible, report results with per-pixel versus regional or national cGDD.
minor comments (5)
  1. [Section 5] The text says 'In each of six cross-year folds, two seasons are held out for testing while the remaining one season is used for training,' but Table 2 lists one training year and one test year per row. This wording should be corrected to describe one training season and one held-out test season per fold.
  2. [Section 4 and Appendix A] The abstract says the dataset is 'publicly released,' but the full text says the dataset and code 'will be publicly available' without a URL or DOI. Provide a permanent link or accession number for the SwissCrop dataset and code.
  3. [Throughout] There are typographical and consistency issues: 'Pastis' versus 'PASTIS,' 'climate becomes extremer,' 'These labels where provided' (should be 'were'), and stray spaces in Table 2 average rows (e.g., '17 .3').
  4. [Figures 7 and 8] The low-data and early-season claims are supported only by figures without numeric values or error bars. Reporting the averaged accuracy and ECE with standard deviations in the text or figure captions would strengthen these secondary claims.
  5. [Section 3.2 and Related Work] The distinction between Thermal-PE [31], which changes positional encodings, and T3S, which changes the sampling grid, is made implicitly through experiments. A short explicit sentence in the method section would help readers understand why the baseline is different in kind.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: T3S's thermal-time sampling rule is fixed a priori from temperature data, its temporal length is set on an external benchmark, and the claimed gains are measured on held-out years against baselines that are not self-cited.

full rationale

The paper's derivation chain is self-contained. GDD in Eq. (1) is computed from daily temperatures with Tbase=0 fixed a priori; cGDD in Eq. (2) is a deterministic accumulation; and Algorithm 1 selects one least-cloudy observation per uniform cGDD bin. None of these ingredients is fitted to crop labels, to the test years, or to the cross-year folds, and no sampling parameter is learned from the evaluation data. The temporal length T=24 is chosen on the independent PASTIS benchmark (Section 3.2, Table 1), not on SwissCrop test folds. Reported accuracy, mIoU, ECE, NLL, and Brier-score gains are measured under leave-one-year-out evaluation on held-out years. The baseline set includes Thermal-PE from Nyborg et al., MC-Dropout, and a Deformable calendar-sampling baseline; the improvement over the Deformable baseline specifically isolates thermal binning from cloud filtering. Self-citations (e.g., [29], [51]) appear in related work and in motivating statements about cloud cover or prior architectures, but the central claim does not rest on any self-cited uniqueness theorem or fitted result. The single-base-temperature assumption (Tbase=0 for all 50 crops) is an ecophysiological modeling choice whose robustness could be questioned, but that is an assumption about the input index, not a circular step, because the method's output is not used to define or fit that index.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The method rests on the ecophysiological premise that GDD captures phenology, on the choice T=24 and Tbase=0, and on the cloud-selection proxy. None of these are fitted to test-year labels, so the circularity burden is low, but the domain assumptions are untested across crop types and regions.

free parameters (2)
  • T (number of sampled timestamps) = 24
    Chosen after PASTIS experiments in Section 3.2 balancing accuracy vs memory; not tuned on SwissCrop test years, but it is a hand-selected hyperparameter that affects all methods equally.
  • Tbase (base temperature for GDD) = 0 °C
    Stated as a common value for temperate crops in Section 3.3; not fitted to data, but no sensitivity analysis is provided.
assumptions (5)
  • domain assumption Crop phenological development is primarily driven by thermal time, i.e., cumulative growing degree days.
    Central premise of T3S; stated in Sections 1 and 3.3, with other drivers (precipitation, management) explicitly out of scope.
  • domain assumption A single base temperature of 0 °C is appropriate for all 50 crop types in SwissCrop.
    Section 3.3 uses Tbase=0 °C as a common temperate value; no per-crop or per-region validation.
  • domain assumption The least-cloudy observation in each equal-cGDD interval is the most informative for classification.
    Algorithm 1 selects argmax of cloud-free pixels; no quantitative check that this selection rule maximizes classification value.
  • domain assumption Temperature at 1 km resolution, resampled to 10 m, is a faithful field-level thermal signal.
    Appendix B describes resampling MeteoSwiss 1 km data to 10 m; no ground-truth temperature validation.
  • domain assumption U-TAE relies on relative ordering rather than absolute timestamps, so re-indexing the sequence is sufficient.
    Section 3.2 PASTIS experiments show linear and calendar positional encodings perform similarly; used to motivate sampling over encoding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of $T^{3}S$: Think in Thermal Time for Generalizable Crop Mapping from Satellite Image Time Series." pith.science (2026). https://pith.science/paper/2DMEWCES

@misc{pith2026250612885,
  author       = {Pith},
  title        = {Pith review of: $T^3S$: Think in Thermal Time for Generalizable Crop Mapping from Satellite Image Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2DMEWCES}},
  note         = {Machine review of arXiv:2506.12885}
}
abstract

Crop type classification from optical satellite time series remains limited in its ability to generalize across growing seasons, particularly when crop phenology shifts due to inter-annual weather variability. This hampers deployment in operational settings where current-year labels are unavailable. In addition, uncertainty quantification is often overlooked, reducing the reliability of such approaches for practical crop monitoring. Inspired by ecophysiological principles, we introduce Thermal Time-based Temporal Sampling ($T^3S$), a simple, model-agnostic method that replaces calendar time with thermal time. By re-indexing satellite observations by cumulative growing degree days, $T^3S$ aligns phenologically equivalent growth stages across years, reducing temporal redundancy while concentrating on the most biologically informative periods. We evaluate $T^3S$ across three architecturally distinct backbones on (i) SwissCrop, a new country-scale, multi-year Sentinel-2 dataset with paired temperature data that we publicly release, and (ii) the cross-region TimeMatch benchmark spanning Denmark and France. Across these settings, $T^3S$ consistently improves cross-year and cross-region crop classification over several state-of-the-art baselines, including thermal positional encoding, with particularly strong gains in uncertainty calibration, robustness under label scarcity, and early-season prediction, while requiring no architectural modification.

Figures

Figures reproduced from arXiv: 2506.12885 by the authors.

Figure 1
Figure 1. Spatially averaged cumulative growing degree days [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. NDVI time series of two parcels with the same crop [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Learning curves of the U-TAE model with different position encodings (PE) on the PASTIS dataset, showing training loss, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Method overview. The first row shows conventional uni [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Class-wise accuracy difference between T3S and U [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: In-season classification performance at 75th- and 50th [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Qualitative comparison between T3S and U-TAE. Green [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Region of interest (top left), the SwissCrop dataset (bot [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: SwissCrop dataset label distribution for the year 2021. Note the logarithmic scale of the y-axis. [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Think with Extra-Image: A Farmland Segmentation Agent Driven by Spatio-Temporal Information Gain

    cs.CV 2026-07 conditional novelty 6.0 of 10

    On-demand spatio-temporal queries guided by ambiguity perception improve high-resolution farmland segmentation stability over single-image and passive multi-image baselines.

Reference graph

Works this paper leans on

65 extracted references · 62 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ai-aristotle: A physics-informed framework for systems biology gray- box identification

    Nazanin Ahmadi Daryakenari, Mario De Florio, Khem- raj Shukla, and George Em Karniadakis. Ai-aristotle: A physics-informed framework for systems biology gray- box identification. PLOS Computational Biology , 20(3): e1011916, 2024. 8

  2. [2]

    Vits for sits: vision transformers for satellite image time series

    S Bai, Z Zhang, M Ding, Y Liu, C Chen, and B Ghanem. Vits for sits: vision transformers for satellite image time series. arXiv preprint, 2023. 3, 13

  3. [3]

    Proba- bilistic crop type mapping for ex-ante modelling and spatial disaggregation

    Josef Baumert, Thomas Heckelei, and Hugo Storm. Proba- bilistic crop type mapping for ex-ante modelling and spatial disaggregation. Ecological Informatics, 83:102836, 2024. 3, 14

  4. [4]

    Country- wide retrieval of forest structure from optical and sar satellite imagery with deep ensembles

    Alexander Becker, Stefania Russo, Stefano Puliti, Nico Lang, Konrad Schindler, and Jan Dirk Wegner. Country- wide retrieval of forest structure from optical and sar satellite imagery with deep ensembles. ISPRS Journal of Photogram- metry and Remote Sensing, 195:269–286, 2023. 14

  5. [5]

    Mapping of crop types and crop sequences with combined time series of sentinel-1, sentinel-2 and landsat 8 data for germany

    Lukas Blickensd ¨orfer, Marcel Schwieder, Dirk Pflugmacher, Claas Nendel, Stefan Erasmi, and Patrick Hostert. Mapping of crop types and crop sequences with combined time series of sentinel-1, sentinel-2 and landsat 8 data for germany. Re- mote sensing of environment, 269:112831, 2022. 3, 13

  6. [6]

    Weight uncertainty in neural networks

    Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. In International Conference on Machine Learning (ICML) , pages 1613–1622, 2015. 14

  7. [7]

    Revisiting the Encoding of Satellite Image Time Series

    Xin Cai, Yaxin Bi, Peter Nicholl, and Roy Sterritt. Revisiting the encoding of satellite image time series. arXiv preprint arXiv:2305.02086, 2023. 3, 13

  8. [8]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794. ACM, 2016. 1

Show all 65 references
  1. [9]

    Multi-year cropland mapping based on remote sensing data: A case study for the khabarovsk territory, russia

    Konstantin Dubrovin, Andrey Verkhoturov, Alexey Stepanov, and Tatiana Aseeva. Multi-year cropland mapping based on remote sensing data: A case study for the khabarovsk territory, russia. Remote Sensing , 16(9):1633,

  2. [10]

    Masksembles for uncertainty estimation

    Nikita Durasov, Andrei Bagrov, and Dmitry Vetrov. Masksembles for uncertainty estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13539–13548, 2021. 3, 14

  3. [11]

    The climate of Switzerland, 2024

    Federal Office of Meteorology and Climatology, Me- teoSwiss. The climate of Switzerland, 2024. 12

  4. [12]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In International Conference on Machine Learning (ICML), pages 1050–1059, 2016. 3, 6, 7, 14

  5. [13]

    Training sample selection for robust multi-year within-season crop classification using machine learning

    Zitian Gao, Danlu Guo, Dongryeol Ryu, and Andrew W Western. Training sample selection for robust multi-year within-season crop classification using machine learning. Computers and Electronics in Agriculture , 210:107927,

  6. [14]

    Panoptic segmentation of satellite image time series with convolu- tional temporal attention networks

    Vivien Sainte Fare Garnot and Loic Landrieu. Panoptic segmentation of satellite image time series with convolu- tional temporal attention networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 4872–4881, 2021. 1, 2, 3, 4, 5, 7, 12, 13

  7. [15]

    Satellite image time series clas- sification with pixel-set encoders and temporal self-attention

    Vivien Sainte Fare Garnot, Loic Landrieu, Sebastien Gior- dano, and Nesrine Chehata. Satellite image time series clas- sification with pixel-set encoders and temporal self-attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12...

  8. [16]

    White, and Michael A

    Cristina G ´omez, Joanne C. White, and Michael A. Wulder. Optical remotely sensed time series data for land cover clas- sification: A review. ISPRS Journal of Photogrammetry and Remote Sensing, 116:55–72, 2016. 1

  9. [17]

    Imbalanced datasets

    Google Developers. Imbalanced datasets. Machine Learning Crash Course, 2025. Accessed June 2025. 4

  10. [18]

    Gustafsson, Martin Danelljan, and Thomas B

    Fredrik K. Gustafsson, Martin Danelljan, and Thomas B. Sch¨on. Evaluating scalable bayesian deep learning methods for robust computer vision. InCVPR Workshops, pages 318– 319, 2019. 14

  11. [19]

    Lora-ensemble: Efficient uncertainty modelling for self-attention networks

    Michelle Halbheer, Dominik J M ¨uhlematter, Alexan- der Becker, Dominik Narnhofer, Helge Aasen, Konrad Schindler, and Mehmet Ozgur Turkoglu. Lora-ensemble: Efficient uncertainty modelling for self-attention networks. arXiv preprint arXiv:2405.14438, 2024. 3, 14

  12. [20]

    Hopcroft, and Kilian Q

    Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Weinberger. Snapshot ensembles: Train 1, get m for free. In International Conference on Learning Representations (ICLR), 2017. 3, 14

  13. [21]

    What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 6, 14

  14. [22]

    General- ization enhancement strategies to enable cross-year cropland mapping with convolutional neural networks trained using historical samples

    Sam Khallaghi, Rahebe Abedi, Hanan Abou Ali, Hamed Alemohammad, Mary Dziedzorm Asipunu, Ismail Alatise, Nguyen Ha, Boka Luo, Cat Mai, Lei Song, et al. General- ization enhancement strategies to enable cross-year cropland mapping with convolutional neural networks trained using...

  15. [23]

    Simple and scalable predictive uncertainty estima- tion using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estima- tion using deep ensembles. In Advances in Neural Informa- tion Processing Systems (NeurIPS), pages 6402–6413, 2017. 3, 14

  16. [24]

    A high-resolution canopy height model of the earth

    Nico Lang, Walter Jetz, Konrad Schindler, and Jan Dirk Wegner. A high-resolution canopy height model of the earth. Nature, 608:259–264, 2022. 14

  17. [25]

    Feature-wise bias amplification

    Klas Leino, Emily Black, Matt Fredrikson, Shayak Sen, and Anupam Datta. Feature-wise bias amplification. arXiv preprint arXiv:1812.08999, 2018. 8

  18. [26]

    Multi-year crop type mapping using sentinel-2 imagery and deep seman- tic segmentation algorithm in the hetao irrigation district in china

    Guang Li, Wenting Han, Yuxin Dong, Xuedong Zhai, Shen- jin Huang, Weitong Ma, Xin Cui, and Yi Wang. Multi-year crop type mapping using sentinel-2 imagery and deep seman- tic segmentation algorithm in the hetao irrigation district in china. Remote Sensing, 15(4):875, 2023. 13

  19. [27]

    Cropsight: Towards a large-scale operational framework for object-based crop type ground truth retrieval using street view and planetscope satellite imagery

    Yin Liu, Chunyuan Diao, Weiye Mei, and Chishan Zhang. Cropsight: Towards a large-scale operational framework for object-based crop type ground truth retrieval using street view and planetscope satellite imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 216:66–89, 20...

  20. [28]

    Refinement of cropland data layer with effective confidence layer interval and image filtering

    Reza Maleki, Falin Wu, Amel Oubara, Loghman Fathollahi, and Gongliu Yang. Refinement of cropland data layer with effective confidence layer interval and image filtering. Agri- culture, 14(8):1285, 2024. 14

  21. [29]

    Crop classification under varying cloud cover with neural ordinary differential equations

    Nando Metzger, Mehmet Ozgur Turkoglu, Stefano D’Aronco, Jan Dirk Wegner, and Konrad Schindler. Crop classification under varying cloud cover with neural ordinary differential equations. IEEE Transactions on Geoscience and Remote Sensing, 60:1–12, 2021. 1, 3, 13

  22. [30]

    The need for biases in learning generaliza- tions

    Tom M Mitchell. The need for biases in learning generaliza- tions. Cognitive Science, 1980. 8

  23. [31]

    Gener- alized classification of satellite image time series with ther- mal positional encoding

    Joachim Nyborg, Charlotte Pelletier, and Ira Assent. Gener- alized classification of satellite image time series with ther- mal positional encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1392–1402, 2022. 3, 6, 7, 13

  24. [32]

    Multi-year mapping of cropping systems in regions with smallholder farms from sentinel-2 images in google earth en- gine

    Hongwei Qi, Ximin Qian, Songhao Shang, and Heyang Wan. Multi-year mapping of cropping systems in regions with smallholder farms from sentinel-2 images in google earth en- gine. GIScience & Remote Sensing, 61(1):2309843, 2024. 3, 13

  25. [33]

    Sitsmamba for crop classification based on satellite image time series.arXiv preprint arXiv:2409.09673, 2024

    Xiaolei Qin, Xin Su, and Liangpei Zhang. Sitsmamba for crop classification based on satellite image time series.arXiv preprint arXiv:2409.09673, 2024. 3, 13

  26. [34]

    Subsam- pling is not magic: Why large batch sizes work for dif- ferentially private stochastic optimisation

    Ossi R ¨ais¨a, Joonas J ¨alk¨o, and Antti Honkela. Subsam- pling is not magic: Why large batch sizes work for dif- ferentially private stochastic optimisation. arXiv preprint arXiv:2402.03990, 2024. 4

  27. [35]

    The eurocropsml time series benchmark dataset for few-shot crop type classification in europe

    Joana Reuss, Jan Macdonald, Simon Becker, Lorenz Richter, and Marco K¨orner. The eurocropsml time series benchmark dataset for few-shot crop type classification in europe. Sci- entific Data, 12(1):664, 2025. 1

  28. [36]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, pa...

  29. [37]

    Dos Santos, Maria Vakalopoulou, Ronny H ¨ansch, Stine Hansen, Keiller Nogueira, Jonathan Prexl, and Devis Tuia

    Ribana Roscher, Marc Russwurm, Caroline Gevaert, Michael Kampffmeyer, Jefersson A. Dos Santos, Maria Vakalopoulou, Ronny H ¨ansch, Stine Hansen, Keiller Nogueira, Jonathan Prexl, and Devis Tuia. Better, not just more: Data-centric machine learning for earth observation. IEEE G...

  30. [38]

    Temporal vegetation modelling using long short-term memory networks for crop identification from medium-resolution multi-spectral satel- lite images

    Marc Rußwurm and Marco K ¨orner. Temporal vegetation modelling using long short-term memory networks for crop identification from medium-resolution multi-spectral satel- lite images. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition Workshops, ...

  31. [39]

    Temporal vegetation modelling using long short-term memory networks for crop identification from medium-resolution multi-spectral satel- lite images

    Marc Rußwurm and Marco Korner. Temporal vegetation modelling using long short-term memory networks for crop identification from medium-resolution multi-spectral satel- lite images. In Proceedings of the IEEE conference on com- puter vision and pattern recognition workshops , p...

  32. [40]

    Multi-temporal land cover classification with sequential recurrent encoders

    Marc Rußwurm and Marco K ¨orner. Multi-temporal land cover classification with sequential recurrent encoders. IS- PRS International Journal of Geo-Information , 7(4):129,

  33. [41]

    Self-attention for raw optical satellite time series classification

    Marc Rußwurm and Marco K ¨orner. Self-attention for raw optical satellite time series classification. ISPRS Journal of Photogrammetry and Remote Sensing , 169:421–435, 2020. 3, 12

  34. [42]

    Breizhcrops: A satellite time series dataset for crop type identification

    Marc Rußwurm, S ´ebastien Lef `evre, and Marco K ¨orner. Breizhcrops: A satellite time series dataset for crop type identification. In ICML Workshop, 2019. 1

  35. [43]

    End-to-end learned early classification of time series for in-season crop type mapping

    Marc Rußwurm, Nicolas Courty, R ´emi Emonet, S ´ebastien Lef`evre, Devis Tuia, and Romain Tavenard. End-to-end learned early classification of time series for in-season crop type mapping. ISPRS Journal of Photogrammetry and Re- mote Sensing, 196:445–456, 2023. 3, 13

  36. [44]

    Lightweight temporal self-attention for classifying satellite images time series

    Vivien Sainte Fare Garnot and Loic Landrieu. Lightweight temporal self-attention for classifying satellite images time series. In International Workshop on Advanced Analytics and Learning on Temporal Data , pages 171–181. Springer,

  37. [45]

    Leverag- ing class hierarchies with metric-guided prototype learning

    Vivien Sainte Fare Garnot and Loic Landrieu. Leverag- ing class hierarchies with metric-guided prototype learning. arXiv preprint arXiv:2007.03047, 2020. 3, 13

  38. [46]

    Eurocrops: The largest harmonized open crop dataset across the european union

    Maja Schneider, Tobias Schelte, Felix Schmitz, and Marco K¨orner. Eurocrops: The largest harmonized open crop dataset across the european union. Scientific Data , 10(1): 612, 2023. 1

  39. [47]

    A sentinel-2 multiyear, multicountry benchmark dataset for crop classification and segmentation with deep learning

    Dimitrios Sykas, Maria Sdraka, Dimitrios Zografakis, and Ioannis Papoutsis. A sentinel-2 multiyear, multicountry benchmark dataset for crop classification and segmentation with deep learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 15...

  40. [48]

    Vits for sits: Vision transformers for satellite image time series

    Michail Tarasiou, Erik Chavez, and Stefanos Zafeiriou. Vits for sits: Vision transformers for satellite image time series. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10418–10428, 2023. 1

  41. [49]

    Attention mechanism-based deep learning approach for wheat yield estimation and uncer- tainty analysis from remotely sensed variables

    Huiren Tian, Pengxin Wang, Kevin Tansey, Jie Wang, Went- ing Quan, and Junming Liu. Attention mechanism-based deep learning approach for wheat yield estimation and uncer- tainty analysis from remotely sensed variables. Agricultural and Forest Meteorology, 356:110183, 2024. 3, 14

  42. [50]

    Deep learning for vegetation clas- sification from optical satellite image time series

    Mehmet Ozgur Turkoglu. Deep learning for vegetation clas- sification from optical satellite image time series. PhD thesis, ETH Zurich, 2023. 1, 12

  43. [51]

    Crop mapping from image time series: deep learning with multi-scale label hierarchies

    Mehmet Ozgur Turkoglu, Stefano D’Aronco, Gregor Perich, Frank Liebisch, Constantin Streit, Konrad Schindler, and Jan Dirk Wegner. Crop mapping from image time series: deep learning with multi-scale label hierarchies. Remote Sensing of Environment, 264, 2021. 1, 3, 12, 13

  44. [52]

    Gating revisited: Deep multi- layer rnns that can be trained

    Mehmet Ozgur Turkoglu, Stefano D’Aronco, Jan Dirk Weg- ner, and Konrad Schindler. Gating revisited: Deep multi- layer rnns that can be trained. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4081–4092, 2021. 13

  45. [53]

    FiLM-Ensemble: Probabilistic deep learning via feature- wise linear modulation

    Mehmet Ozgur Turkoglu, Alexander Becker, H ¨useyin Anil G¨und¨uz, Mina Rezaei, Bernd Bischl, Rodrigo Caye Daudt, 10 Stefano D’Aronco, Jan Dirk Wegner, and Konrad Schindler. FiLM-Ensemble: Probabilistic deep learning via feature- wise linear modulation. In Advances in Neural In...

  46. [54]

    Country-wide cross-year crop map- ping from optical satellite image time series

    Mehmet Ozgur Turkoglu, Helge Aasen, Konrad Schindler, and Jan Dirk Wegner. Country-wide cross-year crop map- ping from optical satellite image time series. Technical re- port, Copernicus Meetings, 2024. 3, 13

  47. [55]

    Hierarchical crop map- ping from satellite image sequences with recurrent neural networks

    Mehmet Ozgur Turkoglu, Stefano D’aranco, Konrad Schindler, and Jan Dirk Wegner. Hierarchical crop map- ping from satellite image sequences with recurrent neural networks. Multitemporal Earth Observation Image Analy- sis: Remote Sensing Image Sequences , page 41, 2024. 3, 13

  48. [56]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 1

  49. [57]

    Equivariance and invariance inductive bias for learning from insufficient data

    Tan Wad, Qianru Sun, Sugiri Pranata, Karlekar Jayashree, and Hanwang Zhang. Equivariance and invariance inductive bias for learning from insufficient data. In European Con- ference on Computer Vision, pages 241–258. Springer, 2022. 8

  50. [58]

    Remote sensing for agricultural applications: A meta-review.Remote sensing of environment, 236:111402, 2020

    Marie Weiss, Fr ´ed´eric Jacob, and Grgory Duveiller. Remote sensing for agricultural applications: A meta-review.Remote sensing of environment, 236:111402, 2020. 1

  51. [59]

    Bayesian learning via stochastic gradient langevin dynamics

    Max Welling and Yee Whye Teh. Bayesian learning via stochastic gradient langevin dynamics. InInternational Con- ference on Machine Learning (ICML), pages 681–688, 2011. 14

  52. [60]

    Segformer: Simple and efficient design for semantic segmentation with transform- ers

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transform- ers. Advances in neural information processing systems, 34: 12077–12090, 2021. 13 11 Appendix A. Training Details...

  53. [61]

    [44] then proposed the lightweight Temporal Atten- 12 Figure 11

    and [15] were the first to apply self-attention to crop classification from Sentinel-2 data and showed improved re- sults. [44] then proposed the lightweight Temporal Atten- 12 Figure 11. SwissCrop dataset label distribution for the year 2021. Note the logarithmic scale of the...

  54. [62]

    [13] investigated ways to optimize the use of multi-year samples in a within-season crop classification model

    demonstrated that reconstructing multi-year Sentinel-2 NDVI time series with Fourier-based fitting and applying Gradient Boosting enables improved crop mapping across different years. [13] investigated ways to optimize the use of multi-year samples in a within-season crop clas...

  55. [63]

    Very recently, [22] shows improved results for cross-year crop mapping combining a U-Net with photometric augmenta- tion, Tversky-Focal loss, and Monte Carlo (MC) Dropout

    benchmarked deep learning segmentation models (e.g., U-Net and SegFormer [60]) for crop mapping across mul- tiple years, while [47] evaluated convLSTM and convS- TAR [52] networks in a similar multi-year context. Very recently, [22] shows improved results for cross-year crop m...

  56. [64]

    In the context of crop yield and type mapping, several re- cent studies have adopted stochastic inference techniques

    extends this ensemble approach to country-scale forest structure estimation at 10m resolution, demonstrating that dense uncertainty fields help flag regions where auxiliary field data are most needed. In the context of crop yield and type mapping, several re- cent studies have...

  57. [65]

    introduces an attention-guided multi-level crop net- work for winter wheat yield estimation at the county scale, applying MC Dropout during inference to derive spatially explicit uncertainty maps of yield forecasts. Similarly, [22] leverages MC Dropout in a U-Net segmentation ...

Pith tools

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