Pith. sign in

REVIEW 3 major objections 4 minor 31 references

Multi-Year Geospatial Reasoning using Interannually-Consistent Historical Predictions as a Free Input Modality

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Feeding a satellite crop-mapping model its own historical predictions as an extra input lifts crop-only F1 by 1.6 percentage points, and a consistently represented external vegetation mask adds about 2.5 more.

desk verdict A genuinely novel way to feed a model its own prediction history, carefully evaluated, but the headline gains depend on an unproven disjointness between the legacy product's training data and this test set. read the letter →

arxiv 2608.05979 v1 pith:LTY4AJCU submitted 2026-08-06 cs.CV

classification cs.CV
keywords geospatialreasoningcrop-typemappingsatelliteimagetimeseriestemporalcontexthistoricalpredictionsEarthobservationCopernicusHRL
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

Operational satellite crop maps are produced every year by the same model, but the archive of that model's own past predictions is almost never fed back into it. This paper argues that a recurring Earth-observation system can use what it already produces: the history of its own class predictions and confidences becomes a free input modality, and an externally supplied vegetation mask is represented consistently inside the model rather than applied only as post-processing. On a pan-European corpus of about 5.4M labelled pixels, adding the prediction history raises crop-only F1 by 1.6 percentage points and corrects a recall-skewed error profile, with the largest gains on perennial and tree crops such as olives, fruits, and nuts. Representing the external mask consistently in both history and target year adds about 2.5 points on the crop classes. The wider claim is that any recurring geospatial or foundation model that emits class maps could adopt the same low-cost recipe.

What carries the argument

The central object is the CTY Embedding Encoder, a branch attached to an existing optical/SAR/meteorology Transformer. Each past prediction is a pair (class code, confidence); the encoder maps the code to a learnable embedding, scales it by the confidence, adds a learnable year-position embedding, and applies a padding-aware Transformer over the year axis followed by masked mean pooling and a projection. A reserved token represents externally masked observations (the BVL), distinguishing no-data, masked, and real crop years. This branch lets the model learn the temporal grammar of land use, including rotations, perennial persistence, and gradual change, directly from its own outputs. The paper shows that raw integer codes fed through the dense or Transformer encoders add nothing, while the embedding encoding is what carries the gain.

What would settle it

Compare the history-based model's gain when the history is generated by a model trained on parcels that are provably disjoint from the test parcels versus a history generated from overlapping training data; if the 1.6- and 2.5-point gains vanish under disjoint training, the reported improvement is partly leakage. A simpler check: retrain the legacy product from scratch on a held-out subset of fields and feed the resulting history, then measure whether the crop-only F1 gain persists.

Watch

Extended reading notes

Core claim

On the paper's own terms, a single-year, single-task pixel classifier can be turned into a multi-year reasoner simply by letting it read its own interannually consistent prediction history and a consistently represented vegetation mask. The author's central claim is that the history branch, implemented as the Crop Type (CTY) Embedding Encoder, is the only single change that clearly helps: mapping each past class code to a confidence-scaled, time-ordered learnable token and attending over the year axis raises crop-only F1 by 1.6 points over the no-history baseline. The gain is not uniform; it concentrates where multi-year reasoning should help, on stable perennials and tree crops (olives +4.6, fruits +3.7, nuts +3.2 points), and it corrects the baseline's systematic over-prediction of crops. When the externally provided Base Vegetation Layer mask is represented consistently in both the target year and the history, crop-only F1 rises another 2.5 points, independent of the inflated no-crop class. The authors evaluate on the 18 crop classes only and report precision and recall separately to keep comparisons fair despite relabelling of non-crop pixels.

Load-bearing premise

The historical predictions fed to the model come from a legacy product, and the paper does not establish that the legacy product's training data were disjoint from the parcels and reference declarations used here for testing, so part of the measured gain could in principle be label leakage rather than genuine multi-year reasoning.

Editorial extensions

If this is right

  • Any recurring Earth-observation product that emits annual or seasonal class maps can feed its own prediction archive back into the model as an input modality, at no labelling cost.
  • Learned multi-year reasoning can replace or simplify rule-based interannual-consistency post-processing, since the model distils the legacy rules and can raise recall where rules cannot.
  • Externally provided ancillary layers should be represented consistently inside the model rather than applied only as an out-of-model mask; the representation choice is worth about 2.5 points on crop classes.
  • History-aware models produce cleaner maps with fewer spurious year-to-year flips on perennial parcels, as shown in the paper's inference tiles.
  • Loss-level rebalancing on top of balanced batch sampling is harmful at continental imbalance ratios; the paper documents a 14.7-point macro-F1 collapse when both are applied.

Reading between the lines

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

  • A natural extension the author leaves implicit is error accumulation: if the history feeding the model is itself model output, systematic errors could reinforce over many operational cycles; the paper's masking and confidence scaling mitigate but do not eliminate this, and quantifying it over many retraining rounds is future work.
  • The same recipe could be attached to a frozen geospatial foundation backbone as a lightweight context branch, since the history encoder is modular and trains jointly; the paper notes foundation models help in label-scarce settings, where history would be an especially cheap signal.
  • A testable extension: evaluate whether the gains persist when the history comes from a model whose training parcels are provably disjoint from the reference labels, which would separate genuine multi-year reasoning from potential leakage through overlapping legacy training data.
  • The per-class precision/recall shifts suggest a practical tweak: gentle oversampling for recall-limited rare classes such as nuts and dry pulses, and the paper's analysis makes that target precise.
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 / 4 minor

Summary. The paper proposes to make a recurring crop-type mapping system reason across years by feeding it two signals it already receives outside the model: the legacy product's own interannually corrected prediction history and the external Base Vegetation Layer (BVL) mask. The history is encoded by a CTY Embedding Encoder (confidence-scaled categorical tokens, year positions, padding-aware Transformer), fused with the existing optical/SAR/meteo backbone; crop-rotation scalars and a soft consistency loss are added. The BVL study compares target-year and history representations of masked pixels. On a pan-European corpus of about 5.4M labelled pixels, the paper reports that adding history raises crop-only F1 by 1.6 percentage points, with the largest gains on perennials (olives +4.6, fruits +3.7, nuts +3.2), and that a consistent BVL representation adds about 2.5 points; the production model reaches 0.86 accuracy and 0.81 crop-only macro-F1.

Significance. If the results hold, the approach is a low-cost, backbone-orthogonal recipe for any recurring Earth-observation product that emits class maps, and it would be one of the few demonstrated uses of a model's own prediction history as an input modality rather than as post-processing. The evaluation design is thoughtful: field-grouped stratified splits, a crop-only metric robust to composition changes, per-class precision/recall reporting, and an informative negative result on inverse-frequency loss weighting. However, the two main numerical claims rest on single-run comparisons and on a history input whose provenance relative to the test labels is not established; these issues must be resolved before the 'free' modality claim is convincing.

major comments (3)
  1. [§3.3 and §5] The paper does not establish that the legacy HRL model's training data are disjoint from the parcels and reference declarations used here for testing. The history input is sampled from the interannual-consistency-fixed HRL product (Signal 1, §3.3), and the field-grouped split in §5 prevents leakage only within the paper's own 5.4M-pixel corpus. Since the reference corpus is pooled from LPIS/GSAA farmer declarations and LUCAS points (§5), and the HRL product is a continental operational service, the legacy model may well have been trained on the same fields or declarations in earlier years. If so, a test pixel's 2017–2023 history would be a compressed form of test-related labels, and the +1.6 pp crop-only F1 gain in Table 3 would be partly label leakage rather than evidence of learned multi-year reasoning. Because the 'free input modality' claim depends on the history being an independent record of past land use, please provide a provenance audit of the HRL training data, or re-evaluate with history restricted to years/parcels the legacy model demonstrably did not train on, or retrain a legacy model on a disjoint training set and check whether the gains persist.
  2. [§6, Tables 1–3 and Fig. 4] All headline numbers are single training runs with no error bars. Table 1 interprets differences as small as 0.34 pp, and the abstract's +1.6 pp and +2.5 pp are point estimates. Without standard deviations across seeds or paired significance tests, these gains could fall within run-to-run noise. Please report mean ± std over at least three seeds for the no-history baseline, the Strategy-C history model, and the two consistent BVL representations, and indicate which differences are statistically reliable.
  3. [§6.4 and Table 2] The attribution of the BVL gain is not clean. The rows of Table 2 vary both the target-year representation and whether masked years are remapped in the history, and the 'ablation' mentioned in the text (masked years left unrepresented in the history) is not shown in the table. To support the claim that the +2.5 pp gain comes from consistently representing the signal in the history, please report the full crossed comparison (target representation × history remapping), including the row with the target representation set to 'dedicated class' or 'folded into no-crop' while the masked years in the history are left unrepresented.
minor comments (4)
  1. [§4.6] The phrase 'soft self-supervised consistency loss' is a misnomer: the auxiliary head is supervised by target labels, or by the main head's soft predictions, which is not self-supervision in the usual sense. Please clarify the target of the consistency loss and rename if appropriate.
  2. [Table 3 and §6.6] Please state explicitly which split (the fixed 720,083-sample 76-source split or the larger 85-source set) is used for each row, since a production model evaluated on the harder 85-source set would not be directly comparable to a baseline evaluated on the 76-source set.
  3. [Fig. 4 and Table 2] The legend 'BVL signal ablated' is confusing given the text says the mask is never removed; consider rewording to 'no special representation'.
  4. [Throughout] There are formatting artifacts such as 'T able 1', 'T able 2', and 'T able 3'; please correct them.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the history is an external legacy product record, the target year is masked, and gains are measured against held-out reference labels; the main caveat is a data-provenance leakage risk, not a definitional reduction.

full rationale

The paper's central claim is that feeding a model its own past predictions, encoded as confidence-scaled categorical tokens, improves crop-type mapping. This is not circular by construction: the history input is sampled from the legacy HRL product (Sec. 3.3), while the target-year label is explicitly masked (Sec. 4.6: 'the current and future years are masked, so the current-year label is never an input'), so the output is not an input in disguise. The soft consistency loss is a weakly weighted regularizer (λ=0.1) attached to a linear head, not a fitted quantity that defines the reported F1 gains; the main loss is focal cross-entropy against external reference labels. Evaluation uses a held-out split that is grouped by field and stratified by class (Sec. 5), and the crop-only metric was introduced precisely to avoid composition artifacts from BVL relabelling. The self-citations present ([5], [25], [26]) are not load-bearing: [5] supports a design choice that is also independently demonstrated by the paper's own 12M-parameter overfitting result (Table 1), and [25] is contextual. The one legitimate concern is data provenance: the history comes from the interannual-fixed product, and the reference corpus is built from LPIS/GSAA and LUCAS sources (Sec. 5); if the legacy HRL model was trained on the same or overlapping parcels, the reported history gain could partly reflect label leakage rather than learned multi-year reasoning. This is a real empirical risk, and the paper's Sec. 7 discussion of error accumulation does not address training-label overlap, but it is not a circularity in the derivation chain: the history is not defined in terms of the target labels, and no parameter is fitted to force the reported improvement. The derivation is therefore self-contained, with the provenance caveat as a correctness risk rather than a circularity finding.

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

No physical or conceptual entities are invented; the CTY embedding and the BVL token are learned model parameters rather than postulated external objects. The central claim rests mainly on data assumptions and hyperparameter choices, listed above.

free parameters (5)
  • consistency loss weight lambda = 0.1
    Set in Sec. 4.6; controls how strongly the soft auxiliary loss enforces agreement between history and target.
  • temperature of consistency head
    Mentioned in Sec. 4.6 but no value or tuning procedure is given.
  • history encoder shape = production: embedding dim 32, 2 layers, 64-d output; main transformer dim 128, 8 heads
    Architecture choices in Sec. 4 and 6.6; the 12M-parameter variant shows results are sensitive to capacity.
  • unclassified probability threshold = 0.35
    Production argmax rule in Sec. 3.1 that determines whether a past prediction is a valid class or unclassified; affects history sampling.
  • history window length = 13 slots, padded to 2029
    Fixed in Sec. 4.3; assumes future years can be zero-padded and that 2017 is always the first available year.
assumptions (5)
  • standard math Transformer self-attention and learned entity embeddings are effective building blocks for this pixel-classification task.
    The method inherits these standard machine-learning primitives without new theoretical justification (Sec. 4.1-4.3).
  • domain assumption The interannual-fixed HRL product history is a faithful record of prior crop state.
    Sec. 3.3 treats the legacy rule-corrected predictions as reliable context; if the legacy product has systematic errors, the history is corrupted.
  • domain assumption The Base Vegetation Layer mask is authoritative and consistently applied across years.
    Sec. 4.5 studies how to read BVL-masked observations but never questions the mask itself.
  • domain assumption Reference labels from LPIS/GSAA and LUCAS are accurate enough to measure crop type.
    Sec. 5 uses farmer declarations and survey points as ground truth without a label-noise analysis.
  • ad hoc to paper Future years can be represented as zero-padded no-data slots without distorting attention.
    Sec. 4.3 pads the 13-slot sequence to 2029; this is a design convenience, not an empirically validated choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Year Geospatial Reasoning using Interannually-Consistent Historical Predictions as a Free Input Modality." pith.science (2026). https://pith.science/paper/LTY4AJCU

@misc{pith2026260805979,
  author       = {Pith},
  title        = {Pith review of: Multi-Year Geospatial Reasoning using Interannually-Consistent Historical Predictions as a Free Input Modality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LTY4AJCU}},
  note         = {Machine review of arXiv:2608.05979}
}
read the original abstract

Machine learning, and deep networks in particular, are increasingly used to derive higher-level Earth observation (EO) products such as annual land-cover and crop-type maps. Many are generated operationally: each year a new acquisition is processed, typically with the same model, extending a multi-year archive. In the process these systems accumulate two kinds of useful signal that are almost never fed back into the model: the system's own archive of past predictions, and ancillary layers produced by other partners in a processing consortium. Both are normally used outside the network, as rule-based post-processing or a fixed input mask. Using the Copernicus Land Monitoring Service High Resolution Layer (HRL) Croplands crop-type product as a testbed, we show that bringing both signals inside the model turns a single-year, single-task pixel classifier into one that reasons across years. We introduce a Crop Type (CTY) embedding encoder that represents each past prediction as a confidence-scaled, time-ordered categorical token and attends over the year axis, and we study how the externally provided Base Vegetation Layer (BVL) mask should be represented in the model's inputs and outputs. To compare designs fairly when they relabel non-crop pixels, we evaluate on the 18 crop classes only and report precision and recall separately. On a pan-European dataset of about 5.4M labelled pixels, adding the prediction history raises crop-only F1 by 1.6 percentage points (pp) and, more importantly, corrects a recall-skewed error profile, with the largest gains on perennial and tree crops (olives +4.6, fruits +3.7, nuts +3.2 pp). Representing the BVL mask consistently in both the history and the target year adds about 2.5 pp on the crop classes. The approach is a low-cost recipe for any recurring geospatial or foundation model that emits class maps.

Figures

Figures reproduced from arXiv: 2608.05979 by the authors.

Figure 1
Figure 1. Operational chain: legacy (top) vs proposed (bottom) Legacy flow Single-year model Rule-based post-processing Interannual consistency fix Product + derived layers EO inputs (S2/S1/meteo) Interannual-fixed history BVL-masked observations History-aware model (learned reasoning) Product (simpler chain) history sampled from the interannual-fixed product [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The multi-source architecture. The green, red, yellow, and purple boxes are the baseline operational model. The blue branch is our addition: two aligned 13-year se￾quences (the product’s own past class predictions and their confidences) are encoded by the CTY Embedding Encoder and concatenated before the MLP head. Crop-rotation scalars enter the dense encoder, and a soft consistency loss is attached to the head. reg… view at source ↗
Figure 3
Figure 3. Reference-data availability across countries (rows) and years (columns). Cov￾erage is uneven across space and time. However, the history input itself is dense as it comes from the prior years’ model predictions, is provided up to year T−1, while the target year T and future years are masked. target at a low weight (λ=0.1). This rewards consistency where history is clearly informative but is too weak to override the … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Crop-only macro precision, recall, and F1 across BVL representations. A con￾sistent representation of masked observations (right two groups) improves the crop classes over giving them no special representation (left), independent of the inflated no-crop class. equalise…
Figure 5
Figure 5. Figure 5: Effect of adding prediction history, per crop class. (a) Precision/recall shift from the no-history model (open markers) to the history model (filled); arrows show the movement. Most classes move toward the precision-equals-recall diagonal while also moving outward to …
Figure 6
Figure 6. Figure 6: Inference tile sample from 2024 production model (no interannual layer) vs. from the proposed model: fewer spurious year-to-year flips and smoother field interiors. History is itself model output, so errors could in principle reinforce them￾selves, and a model could le…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 21 canonical work pages

  1. [1]

    IEEE Trans

    Abercrombie, S.P., Friedl, M.A.: Improving the consistency of multitemporal land cover maps using a hidden Markov model. IEEE Trans. Geosci. Remote Sens. 54(2), 703–713 (2016).https://doi.org/10.1109/TGRS.2015.2463689

  2. [2]

    In: IEEE/CVF Conf

    Astruc, G., Gonthier, N., Mallet, C., Landrieu, L.: AnySat: One earth observation model for many resolutions, scales, and modalities. In: IEEE/CVF Conf. Comput. Vis. Pattern Recog. (CVPR) (2025)

  3. [3]

    In: IEEE Int

    Bailly, S., Giordano, S., Landrieu, L., Chehata, N.: Crop-rotation structured clas- sification using multi-source Sentinel images and LPIS for crop type mapping. In: IEEE Int. Geosci. Remote Sens. Symp. (IGARSS) (2018).https://doi.org/10. 1109/IGARSS.2018.8518427

  4. [4]

    Scientific Data9, 251 (2022).https://doi.org/10.1038/s41597-022-01307-4

    Brown, C.F., Brumby, S.P., Guzder-Williams, B., Birch, T., Hyde, S.B., Maz- zariello, J., Czerwinski, W., Pasquarella, V.J., Haertel, R., Ilyushchenko, S., et al.: Dynamic world, near real-time global 10m land use land cover mapping. Scientific Data9, 251 (2022).https://doi.org/10.1038/s41597-022-01307-4

  5. [5]

    In: Proc

    Butsko, C., Tricht, K.V., Tseng, G., Milli, G., Rolnick, D., Cartuyvels, R., Becker- Reshef, I., Szantoi, Z., Kerner, H.: Deploying geospatial foundation models in the real world: Lessons from WorldCereal. In: Proc. TerraBytes ICML Workshop: To- wards Global Datasets and Models for Earth Observation. Proc. Mach. Learn. Res. (PMLR), vol. 292, pp. 13–31 (2025)

  6. [6]

    In: NeurIPS (2022)

    Cong, Y., Khanna, S., Meng, C., Liu, P., Rozi, E., He, Y., Burke, M., Lobell, D.B., Ermon, S.: SatMAE: Pre-training transformers for temporal and multi-spectral satellite imagery. In: NeurIPS (2022)

  7. [7]

    Document Code D1.12, CLMS (2025)

    Copernicus Land Monitoring Service: HRL croplands product user manual, version 2.3 (cty). Document Code D1.12, CLMS (2025)

  8. [8]

    Document Code D1.2, CLMS (2025)

    Copernicus Land Monitoring Service: HRL VLCC algorithm theoretical basis doc- ument (atbd), version 2.2. Document Code D1.2, CLMS (2025)

Show all 31 references
  1. [9]

    In: CVPR (2019)

    Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: CVPR (2019)

  2. [10]

    Remote Sensing of Environment266, 112416 (2021).https://doi.org/10.1016/j.rse.2021

    d’Andrimont, R., Verhegghen, A., Meroni, M., Lemoine, G., Defourny, P., van der Velde, M.: From parcel to continental scale – a first european crop type map based on Sentinel-1 and LUCAS Copernicus in-situ observations. Remote Sensing of Environment266, 112416 (2021).https://d...

  3. [11]

    arXiv preprint arXiv:2506.06281 (2025)

    Danish, M.S., Munir, M.A., Shah, S.R.A., Khan, M.H., Anwer, R.M., Laaksonen, J., Khan, F.S., Khan, S.: TerraFM: A scalable foundation model for unified multi- sensor earth observation. arXiv preprint arXiv:2506.06281 (2025)

  4. [12]

    In: Advanced Analytics and Learning on Temporal Data (AALTD), ECML PKDD Workshop (2020) 16 S.R.Ali Shah et al

    Garnot, V.S.F., Landrieu, L.: Lightweight temporal self-attention for classifying satellite image time series. In: Advanced Analytics and Learning on Temporal Data (AALTD), ECML PKDD Workshop (2020) 16 S.R.Ali Shah et al

  5. [13]

    In: ICCV (2021)

    Garnot, V.S.F., Landrieu, L.: Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In: ICCV (2021)

  6. [14]

    Photogrammetric En- gineering & Remote Sensing86(7), 431–441 (2020).https://doi.org/10.14358/ PERS.86.7.431

    Giordano, S., Bailly, S., Landrieu, L., Chehata, N.: Improved crop classification with rotation knowledge using Sentinel-1 and -2 time series. Photogrammetric En- gineering & Remote Sensing86(7), 431–441 (2020).https://doi.org/10.14358/ PERS.86.7.431

  7. [15]

    arXiv preprint arXiv:1604.06737 (2016)

    Guo, C., Berkhahn, F.: Entity embeddings of categorical variables. arXiv preprint arXiv:1604.06737 (2016)

  8. [16]

    In: NeurIPS Deep Learning Workshop (2015)

    Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. In: NeurIPS Deep Learning Workshop (2015)

  9. [17]

    In: arXiv preprint arXiv:2310.18660 (2023)

    Jakubik, J., Roy, S., Phillips, C.E., Fraccaro, P., Godwin, D., et al.: Founda- tion models for generalist geospatial artificial intelligence. In: arXiv preprint arXiv:2310.18660 (2023)

  10. [18]

    In: ICCV (2017)

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object detection. In: ICCV (2017)

  11. [19]

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Int. Conf. Learn. Represent. (ICLR) (2019)

  12. [20]

    Remote Sensing11(5), 523 (2019)

    Pelletier, C., Webb, G.I., Petitjean, F.: Temporal convolutional neural network for the classification of satellite image time series. Remote Sensing11(5), 523 (2019). https://doi.org/10.3390/rs11050523

  13. [21]

    Remote Sensing13(22), 4599 (2021)

    Quinton, F., Landrieu, L.: Crop rotation modeling for deep learning-based par- cel classification from satellite time series. Remote Sensing13(22), 4599 (2021). https://doi.org/10.3390/rs13224599

  14. [22]

    ISPRS Journal of Photogrammetry and Remote Sensing169, 421–435 (2020).https://doi.org/10.1016/j.isprsjprs.2020.06.006

    Rußwurm, M., Körner, M.: Self-attention for raw optical satellite time series clas- sification. ISPRS Journal of Photogrammetry and Remote Sensing169, 421–435 (2020).https://doi.org/10.1016/j.isprsjprs.2020.06.006

  15. [23]

    In: ISPRS Congress (2020)

    Rußwurm, M., Pelletier, C., Zollner, M., Lefèvre, S., Körner, M.: BreizhCrops: A time series dataset for crop type mapping. In: ISPRS Congress (2020)

  16. [24]

    Scientific Data10, 612 (2023)

    Schneider, M., Schelte, T., Schmitz, F., Körner, M.: EuroCrops: A pan-european dataset for time series crop type classification. Scientific Data10, 612 (2023). https://doi.org/10.1038/s41597-023-02517-0

  17. [25]

    Journal of Spatial Science69(3), 849–872 (2024)

    Shah, S.R.A., Obaid-Ur-Rehman, Shabbir, Y., Ishaq, R.A.: Contextual band ad- dition and multi-look inferencing to improve semantic segmentation model per- formance on satellite images. Journal of Spatial Science69(3), 849–872 (2024). https://doi.org/10.1080/14498596.2024.2305124

  18. [26]

    Earth System Science Data15, 5491–5515 (2023)

    Tricht, K.V., Degerickx, J., Gilliams, S., Zanaga, D., Battude, M., et al.: World- Cereal: a dynamic open-source system for global-scale, seasonal, and reproducible crop and irrigation mapping. Earth System Science Data15, 5491–5515 (2023). https://doi.org/10.5194/essd-15-5491-2023

  19. [27]

    In: NeurIPS Workshop on Tackling Climate Change with Machine Learning (2023)

    Tseng, G., Cartuyvels, R., Zvonkov, I., Purohit, M., Rolnick, D., Kerner, H.: Lightweight, pre-trained transformers for remote sensing timeseries. In: NeurIPS Workshop on Tackling Climate Change with Machine Learning (2023)

  20. [28]

    In: NeurIPS (2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, L., Polosukhin, I.: Attention is all you need. In: NeurIPS (2017)

  21. [29]

    Wang, Y., Feng, L., Zhang, Z., Tian, F.: An unsupervised domain adaptation deep learning method for spatial and temporal transferable crop type mapping using Sentinel-2 imagery. ISPRS J. Photogramm. Remote Sens.199, 102–117 (2023)

  22. [30]

    Yuan, Y., Lin, L.: Self-supervised pretraining of transformers for satellite image time series classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.14, 474–487 (2021).https://doi.org/10.1109/JSTARS.2020.3036602 Historical Predictions as a Free Input Modality 17

  23. [31]

    In: Proc

    Zerveas, G., Jayaraman, S., Patel, D., Bhamidipaty, A., Eickhoff, C.: A transformer-based framework for multivariate time series representation learning. In: Proc. 27th ACM SIGKDD Conf. Knowledge Discovery and Data Mining (KDD) (2021).https://doi.org/10.1145/3447548.3467401

Pith tools

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