Pith. sign in

REVIEW 3 major objections 4 minor 69 references

Farm-Level, In-Season Crop Identification for India

T0 review · 3 major / 4 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 satellite-driven deep learning system can identify 12 major crops at farm level across India within the growing season, matching the official 2023-24 census at 94% (winter) and 75% (monsoon) cosine similarity on state-level areas.

desk verdict A credible, important engineering effort whose headline census numbers rely on an unvalidated season-detection algorithm and a weak cosine metric; deserves peer review with requests for validation. read the letter →

arxiv 2507.02972 v1 pith:MJUCDG4Y submitted 2025-06-30 cs.CV cs.LG

Ishan Deshpande , Amandeep Kaur Reehal , Chandan Nath , Renu Singh , Aayush Patel , Aishwarya Jayagopal , Gaurav Singh , Gaurav Aggarwal
show 15 more authors
Amit Agarwal Prathmesh Bele Sridhar Reddy Tanya Warrier Kinjal Singh Ashish Tendulkar Luis Pazos Outon Nikita Saxena Agata Dondzik Dinesh Tewari Shruti Garg Avneet Singh Harsh Dhand Vaibhav Rajan Alok Talekar
This is my paper
classification cs.CVcs.LG
keywords croptypemappingin-seasonidentificationSentinel-1Sentinel-2seasondetectionfarmboundariestransformercensusvalidation
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 satellite-based deep learning system can identify which of 12 major crops is growing in each farm field across India while the season is still underway, using Sentinel-1 and Sentinel-2 imagery combined with national farm-boundary data. The twelve crops cover nearly 90% of India's cultivated area, and when the model's predicted crop areas are summed per state they agree with the official 2023-24 agricultural census with cosine similarity 0.94 in the winter season and 0.75 in the monsoon. The system also estimates each field's sowing and harvest dates automatically, so predictions become reliable about two months into the growing season and can be re-issued monthly. If correct, India would have what the authors describe as its first pan-India, in-season, farm-level crop map, providing a data source for early production estimates, food security planning, and insurance decisions that previously had to wait for post-season surveys.

What carries the argument

The load-bearing mechanism is the two-stage pipeline that converts raw field labels into season-relative, in-season training examples. First, the season detection algorithm (from [32]) turns a bare/vegetated NDVI time series, thresholded after cloud masking and smoothing, into estimated sowing and harvest dates per field; these dates filter which labels are usable, define the 'days after estimated season start' axis along which the model is trained and evaluated, and assign seasons for the census comparison. Second, the classifier is a transformer encoder with separate attention layers per satellite (pre-fusion encoders) and shared layers across satellites (post-fusion encoders), pre-trained with a masked-autoencoder objective on unlabeled Indian data and fine-tuned with focal loss to predict the active crop from one year of interpolated, field-median Sentinel-1/2 signal ending at the prediction date. Field-level aggregation uses ALU farm boundaries, which turn point labels into farm units and define the area weights in the census comparison.

What would settle it

Take a sample of fields in one or two states, obtain independent field-verified sowing dates for one season from agricultural extension records or a fresh survey, and compare the model's predictions at estimated day 60 against predictions at true day 60 measured from the verified dates. If precision and recall at true day 60 are substantially below the reported day-60 numbers, the 'two months into the season' capability largely reflects a late-biased season-start estimate rather than genuine early-season skill.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that in-season, farm-level crop identification at national scale is achievable: a transformer model trained on roughly 70,000 field-verified labels, augmented to about 208,000 season-sliced examples, can classify 12 crops plus an Others bucket, with per-crop precision and recall that rise over the first 60-90 days after estimated season start and aggregate to state-level area distributions matching the 2023-24 census (0.94 winter, 0.75 monsoon cosine similarity across 26 states). The paper further claims this is enabled by a two-stage design: an automated season detection algorithm that estimates each field's sowing and harvest dates, and a classifier that only sees one year of satellite signal ending at the prediction date. A key design choice is excluding location and absolute time from the model inputs, relying only on the satellite signal so the model generalises across regions; the reported early-season skill (non-zero performance at day 0) is attributed to the season detector lagging actual sowing slightly and to the model exploiting common crop-rotation patterns.

Load-bearing premise

The entire system trusts the season detection algorithm, which has appeared only as an extended abstract, to estimate each field's sowing and harvest dates correctly; those estimates filter the training data, define what 'two months into the season' means, and assign every field to a season for the census comparison, so a systematic bias in the dates would distort both the in-season claim and the reported census agreement.

Editorial extensions

If this is right

  • Crop-area estimates for wheat, mustard, rice, and other major crops become available roughly two months after sowing and refresh monthly, supporting in-season production forecasts and procurement planning before harvest.
  • State-level predicted areas can be checked directly against official statistics; the 0.94 winter agreement means the product could serve as an independent, higher-cadence complement to annual census reporting.
  • Because the classifier receives no location or absolute-time features, its learned crop signatures should transfer across regions within India and to other smallholder farming countries, so extending coverage to more crops or new geographies requires only new labels, not new model design.
  • The monsoon gap (0.75) is the stated frontier; the paper attributes it to weaker early-season precision, notably cotton overprediction, making near-term improvement in monsoon discrimination a concrete, testable target.
  • A monthly, per-state area series like the one presented can be built operationally, turning the system into an early-warning signal for crop anomalies rather than a post-season map.

Reading between the lines

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

  • My inference: the 'two months into the growing season' claim is only as strong as the season-start estimates; if the detector systematically dates sowing late, day-60 predictions are effectively later in real phenological time, and measuring true sowing dates on a held-out set of fields would settle how much of the early-season skill is real.
  • My inference: since training labels come from 7 states but evaluation covers 26, the reported census agreement is an implicit test of geographic generalisation; a per-state stratified analysis of cosine similarity, which the paper's table already permits, could reveal which agro-climatic zones drive the monsoon weakness.
  • My inference: the static-boundary assumption on ALU fields means area totals drift as real field boundaries change; pairing the system with an updated boundary layer each season would tighten the census comparison, especially for crops like cotton.
  • My inference: the same architecture and season-relative slicing could be re-pointed at additional crops, such as horticultural or mixed-cropping systems, wherever a season detector and field boundaries exist, making the 12-crop ceiling a data constraint rather than a method constraint.
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. This paper presents a deep learning pipeline for farm-level, in-season crop type identification across India, using Sentinel-1/2 time series aggregated over ALU field boundaries and a transformer encoder-decoder pre-trained with masked autoencoding. A season detection algorithm from an unpublished AGU abstract (Ref. [32]) estimates sowing and harvest dates, and is used to filter training samples, define the in-season evaluation clock, and assign predictions to winter/monsoon seasons for comparison with the 2023-24 national crop census. The authors report field-level precision/recall that improve through the season, and state-level cosine similarities of 0.94 (winter) and 0.75 (monsoon) between predicted and census areas.

Significance. If the claims hold, this would be a valuable contribution to agricultural monitoring, providing a national-scale in-season crop map with farm-level resolution. The paper has several strengths: it addresses a large and practically important problem; it uses publicly available multi-sensor satellite data; it employs a plausible model architecture with pre-training and data augmentation; it reports field-level metrics with means and standard deviations over three splits; and it attempts a large-scale comparison against official census data. However, the central empirical claims rest on an unvalidated season detection component and a weak aggregate evaluation metric, which limits the current evidence for the headline numbers.

major comments (3)
  1. [Secs. 3.2.2, 4.2, Appendix E] The season detection algorithm [32] is load-bearing at three points: it filters 48,285 labeled fields down to 32,208 (Sec. 3.2.2, step 3), it defines the in-season evaluation clock 'days after estimated season start' (Sec. 4.2), and it assigns every field prediction to Winter or Monsoon for the census comparison (Appendix E, step 5). The authors use [32] without modification and cite only an AGU abstract; no validation of the estimated sowing/harvest dates is provided. The algorithm itself (Appendix B) relies on arbitrary thresholds (NDVI >= 0.4 after a 20-day moving average, CloudScore+ cs_cdf < 0.6, season-merge gap of 1 month) with no sensitivity analysis. The paper's own Sec. 4.3 admits that 'cotton predictions for winter which is atypical ... may be due to incorrect season detection or improper season attribution,' which is direct internal evidence that season-attribution errors reach the census comparison. Because all three headline claims (the training filter, the 'two months into the growing season' result, and the 0.94/0.75 census agreement) depend on this component, the authors should validate [32] against ground-truth sowing/harvest dates for at least a subset of fields, or provide a sensitivity analysis showing the results are robust to plausible errors in the estimated season start.
  2. [Sec. 4.3, Eq. (6), Table 2] The census agreement is measured by cosine similarity between state-level predicted and census area vectors (Eq. 6). This metric is insensitive to overall area under/over-estimation: a prediction vector that is a constant multiple of the census vector gives cosine similarity 1.0. The OVERALL row in Table 2 (0.94 winter, 0.75 monsoon) is computed on the aggregate national vectors, which can mask severe failures in smaller states (e.g., Assam monsoon 0.23, Himachal Pradesh 0.20, Nagaland 0.21). Moreover, Sec. 4.3 states that the comparison is for 'the major crops in each state,' but the exact crop set used per state is not defined; if only 4-5 crops are compared, the cosine similarity is computed over a very small vector and may be high by chance. The authors should report additional aggregate metrics (e.g., per-crop mean absolute percentage error, area-weighted per-state cosine similarity, or a confusion matrix at state level) and specify the exact crop set for each state.
  3. [Secs. 3.2.2-3.2.3, 4.2] The field-level evaluation only uses test examples that survived the season-detection filter of Sec. 3.2.2 step 3. Because the same filter is applied to the training and test data, the reported precision/recall (Figs. 4-5, Appendix F) may not be representative of all fields, especially those for which season detection fails. Additionally, the ground-truth labels are collected from only 7 states (Appendix A.3), so the field-level metrics are not pan-Indian; the census evaluation covers more states but is based on a different, aggregate comparison. The authors should acknowledge this limitation and, ideally, evaluate on an independent set of fields with known crop labels that have not been filtered by season detection.
minor comments (4)
  1. [Abstract and Sec. 1] The claim that the 12 crops 'collectively account for nearly 90% of India's total cultivated area' is not accompanied by a citation; the source of this statistic should be given.
  2. [Appendix D] Table D3 lists hyper-parameters explored, but the paper does not report the final values chosen for the deployed models; providing the selected configuration would improve reproducibility.
  3. [Sec. 4.2] The assignment of test examples to Winter/Monsoon by label timestamp (June-October vs November-March) is a coarse heuristic that may misclassify long-duration or multi-season crops such as sugarcane; the authors should note this limitation or justify the partition.
  4. [Tables F4 and F5] Several precision values are 0.00 or NaN for multiple crops in later time steps (e.g., winter sorghum at 180 days, monsoon chilli and gram at all time steps); the authors should comment on whether this reflects lack of test samples, model failure, or season attribution artifacts.

Circularity Check

2 steps flagged · score 6.0 of 10

The in-season timing claim and the Winter/Monsoon split of the census comparison are defined by the authors' own unvalidated season-detection algorithm [32], making the headline 'two months into the growing season' partially circular.

  1. self citation load bearing [Sec. 3, Sec. 3.2.2 Step 3, Appendix B; reference [32]]
    "For the first stage, we use the season detection algorithm from [32] without modifications. ... We use the season detection algorithm from [32] to identify the crop season bounds associated with each sample, defined as (estimated season start time, estimated season end time)."

    Reference [32] is an AGU abstract by the same author group (Kaur, Patel, Deshpande, Nath, Singh, Singh, Talekar) and is not peer-reviewed in full or released. This algorithm is load-bearing at three points: it filters the training data from 48,285 to 32,208 samples (Sec. 3.2.2 Step 3), it defines every in-season evaluation timestamp as 'days after estimated season start' (Sec. 4.2), and it assigns each field to Winter or Monsoon for the census cosine-similarity comparison (Appendix E Step 5). No independent validation of the estimated sowing/harvest dates is provided, so the central in-season claim rests on a self-citation chain rather than on an externally verified component.

  2. self definitional [Sec. 3.2.2 Step 4 and Sec. 4.2]
    "In this paper we refer to the difference between tend and the estimated season start as days after estimated season start. ... Note that this is only an estimate as we are relying on the season detection algorithm for getting the season start/end dates."

    The evaluation timeline of the in-season claim is defined relative to the same estimated season start that the season-detection algorithm produces and that was used to construct the training windows (Eq. 2). Therefore the claim 'reliable crop identification as early as two months into the growing season' is measured against the algorithm's own estimate of the growing season, not against independent sowing/harvest ground truth. If the season-start estimate is biased, the 'two months' claim is biased in the same direction, so this evaluation cannot independently validate the season-detection component or the true in-season timing.

full rationale

The crop classifier itself is trained on 32,208 labeled fields with location-based held-out splits and is compared against external census data, so the core crop-identification result is not circular. However, the paper's central in-season claim is constructed around the season-detection algorithm from [32], which is an unpublished AGU abstract by the same author group. The same algorithm filters the training data (Sec. 3.2.2 Step 3), defines every evaluation timestamp as 'days after estimated season start' (Sec. 3.2.2 Step 4, Sec. 4.2), and assigns each field to Winter/Monsoon for the census cosine-similarity comparison (Appendix E Step 5). The paper explicitly concedes that this is 'only an estimate' and that atypical winter cotton predictions 'may be due to incorrect season detection or improper season attribution.' Because no independent ground-truth sowing/harvest dates or external validation of [32] are provided, the 'two months into the growing season' claim and the 94%/75% seasonal split are measured relative to the algorithm's own estimates, so those headline numbers cannot falsify the season-detection component. This is a partial circularity, not a full one: the crop labels and census totals still provide independent signal, but the seasonal and in-season framing is self-referential.

Assumptions & free parameters 9 free parameters · 6 assumptions · 0 invented entities

The central claim depends on several components that are either proprietary (ALU boundaries, licensed ground truth) or self-cited (season detection [32]). The per-crop season-length filters and NDVI thresholds are hand-chosen and affect both training and evaluation. No new physical entities are introduced.

free parameters (9)
  • NDVI threshold = 0.4
    Appendix B.2 step 5: field declared vegetated when NDVI >= 0.4; used in season detection algorithm from [32].
  • Moving average window = 20 days
    Appendix B.2 step 4: smoothened NDVI with 20-day moving average before thresholding.
  • CloudScore+ cs_cdf threshold = 0.6
    Appendix B.2 step 3: NDVI values with cs_cdf < 0.6 dropped as cloudy.
  • Season merge gap = 1 month
    Appendix B.2 step 7: two detected seasons merged if gap less than 1 month.
  • Per-crop max season length = wheat 240, sugarcane 600, soybeans 150, mustard 150, corn 180, rice 200, cotton 300, gram 180, sorghum 180, groundnut…
    Appendix C, Table C2: samples whose detected season exceeds these bounds are discarded from training, biasing data toward normative phenology.
  • Temporal augmentation interval = 30 days
    Sec 3.2.2 step 4: tend sampled at 30-day intervals from season start to end to create augmented samples.
  • Lookback window = 1 year
    Sec 3.2.2 step 5: each in-season example uses exactly one year of RSD before tend.
  • Focal loss gamma = not stated (0, 1.5, 2.0 explored)
    Appendix D: hyperparameters explored but final selected values not reported in paper.
  • Transformer hyperparameters = not stated (token dim 16 or 64, layers 2-6, etc.)
    Appendix D Table D3 lists explored values; final configuration not specified, hampering reproduction.
assumptions (6)
  • domain assumption ALU farm boundary predictions are accurate and time-invariant
    Sec 4.3: 'we assume the ALU boundaries are constant indefinitely, which might not be true'; boundaries from [41], a Google model not publicly released.
  • domain assumption Ground truth survey labels are correct
    Appendix A.1: labels collected by surveyors physically visiting fields; no independent verification or error rate reported.
  • domain assumption Season detection estimates accurate sowing/harvest dates
    Paper relies on [32] season detection throughout training, in-season evaluation, and census comparison; algorithm not independently validated in this paper.
  • domain assumption State-level crop census 2023-24 is an unbiased ground truth for area
    Sec 4.3 uses census [49] as reference for cosine similarity; census itself has known reporting limitations, and no uncertainty is modeled.
  • domain assumption Median aggregation and 5-day interpolation preserve crop signal
    Sec 3.2.2 step 2: field-level RSD constructed by temporal interpolation to 5-day steps and median over interior points, following [36,37].
  • domain assumption A crop's spectral signature generalizes across India
    Model trained on labels from 7 states and applied to all 26 states; no location features are provided, relying on spectral generalization (Sec 3.2.2 step 6).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Farm-Level, In-Season Crop Identification for India." pith.science (2026). https://pith.science/paper/MJUCDG4Y

@misc{pith2026250702972,
  author       = {Pith},
  title        = {Pith review of: Farm-Level, In-Season Crop Identification for India},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJUCDG4Y}},
  note         = {Machine review of arXiv:2507.02972}
}
read the original abstract

Accurate, timely, and farm-level crop type information is paramount for national food security, agricultural policy formulation, and economic planning, particularly in agriculturally significant nations like India. While remote sensing and machine learning have become vital tools for crop monitoring, existing approaches often grapple with challenges such as limited geographical scalability, restricted crop type coverage, the complexities of mixed-pixel and heterogeneous landscapes, and crucially, the robust in-season identification essential for proactive decision-making. We present a framework designed to address the critical data gaps for targeted data driven decision making which generates farm-level, in-season, multi-crop identification at national scale (India) using deep learning. Our methodology leverages the strengths of Sentinel-1 and Sentinel-2 satellite imagery, integrated with national-scale farm boundary data. The model successfully identifies 12 major crops (which collectively account for nearly 90% of India's total cultivated area showing an agreement with national crop census 2023-24 of 94% in winter, and 75% in monsoon season). Our approach incorporates an automated season detection algorithm, which estimates crop sowing and harvest periods. This allows for reliable crop identification as early as two months into the growing season and facilitates rigorous in-season performance evaluation. Furthermore, we have engineered a highly scalable inference pipeline, culminating in what is, to our knowledge, the first pan-India, in-season, farm-level crop type data product. The system's effectiveness and scalability are demonstrated through robust validation against national agricultural statistics, showcasing its potential to deliver actionable, data-driven insights for transformative agricultural monitoring and management across India.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 56 canonical work pages

  1. [32]

    In: AGU Fall Meeting Abstracts, vol

    Kaur, A., Patel, A., Deshpande, I., Nath, C., Singh, G., Singh, R., Talekar, A.: Agricultural Season Detection At Scale. In: AGU Fall Meeting Abstracts, vol. 2023, pp. 42–07 (2023)

    work page 2023

  2. [1]

    Indian agriculture towards 2030, 1–7 (2022)

    Chand, R.: Indian agriculture towards 2030—need for a transformative vision. Indian agriculture towards 2030, 1–7 (2022)

    work page 2022

  3. [2]

    Food and Agriculture Organization (2020)

    F AO: F AO in India, India at a glance. Food and Agriculture Organization (2020)

    work page 2020

  4. [3]

    Computers and Electronics in Agriculture 210, 107927 (2023)

    Gao, Z., Guo, D., Ryu, D., Western, A.W.: Training sample selection for robust multi-year within-season crop classification using machine learning. Computers and Electronics in Agriculture 210, 107927 (2023)

    work page 2023

  5. [4]

    ISPRS Journal of Photogrammetry and Remote Sensing 194, 250–266 (2022)

    Zhao, L., Li, Q., Chang, Q., Shang, J., Du, X., Liu, J., Dong, T.: In-season crop type identification using optimal feature knowledge graph. ISPRS Journal of Photogrammetry and Remote Sensing 194, 250–266 (2022)

    work page 2022

  6. [5]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Soler, J.L., Friedel, T., Wang, S.: Combining deep learning and street view imagery to map smallholder crop types. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 22202–22212 (2024)

    work page 2024

  7. [6]

    Remote Sensing 14(12), 2758 (2022)

    Yao, J., Wu, J., Xiao, C., Zhang, Z., Li, J.: The classification method study of crops remote sensing with deep learning, machine learning, and google earth engine. Remote Sensing 14(12), 2758 (2022)

    work page 2022

  8. [7]

    Agricultural Systems 201, 103462 (2022)

    Zhang, C., Di, L., Lin, L., Li, H., Guo, L., Yang, Z., Yu, E.G., Di, Y., Yang, A.: Towards automation of in-season crop type mapping using spatiotemporal crop information and remote sensing data. Agricultural Systems 201, 103462 (2022)

    work page 2022

Show all 69 references
  1. [8]

    Remote Sensing 13(7), 1394 (2021)

    Xie, Q., Lai, K., Wang, J., Lopez-Sanchez, J.M., Shang, J., Liao, C., Zhu, J., Fu, H., Peng, X.: Crop monitoring and classification using polarimetric radarsat- 2 time-series data across growing season: A case study in southwestern ontario, canada. Remote Sensing 13(7), 1394 (2021)

    2021

  2. [9]

    ISPRS Journal of Photogrammetry and Remote Sensing 161, 109–123 (2020)

    You, N., Dong, J.: Examining earliest identifiable timing of crops using all available sentinel 1/2 imagery and google earth engine. ISPRS Journal of Photogrammetry and Remote Sensing 161, 109–123 (2020)

    2020

  3. [10]

    https: //arxiv.org/abs/2009.10189

    Kerner, H., Sahajpal, R., Skakun, S., Becker-Reshef, I., Barker, B., Hosseini, M., Puricelli, E., Gray, P.: Resilient In-Season Crop Type Classification in Multi- spectral Satellite Observations using Growth Stage Normalization (2020). https: //arxiv.org/abs/2009.10189

    2020 arXiv

  4. [11]

    Remote Sensing of Environment 274, 112994 (2022)

    Lin, C., Zhong, L., Song, X.-P., Dong, J., Lobell, D.B., Jin, Z.: Early-and in- season crop type mapping without current-year ground truth: Generating labels from historical information via a topology-based approach. Remote Sensing of Environment 274, 112994 (2022)

    2022

  5. [12]

    Frontiers in Plant Science 14 (2023)

    Wang, H., Chang, W., Yao, Y., Yao, Z., Zhao, Y., Li, S., Liu, Z., Zhang, X.: 20 Cropformer: A new generalized deep learning classification approach for multi- scenario crop classification. Frontiers in Plant Science 14 (2023)

    2023

  6. [13]

    Remote Sensing 15(3), 799 (2023)

    Weilandt, F., Behling, R., Goncalves, R., Madadi, A., Richter, L., Sanona, T., Spengler, D., Welsch, J.: Early crop classification via multi-modal satellite data fusion and temporal attention. Remote Sensing 15(3), 799 (2023)

    2023

  7. [14]

    https://arxiv.org/abs/2304.14065

    Tseng, G., Cartuyvels, R., Zvonkov, I., Purohit, M., Rolnick, D., Kerner, H.: Lightweight, Pre-trained Transformers for Remote Sensing Timeseries (2024). https://arxiv.org/abs/2304.14065

    2024 arXiv

  8. [15]

    Agronomy 13(7) (2023) https://doi.org/10.3390/agronomy13071723

    Mao, M., Zhao, H., Tang, G., Ren, J.: In-season crop type detection by combing sentinel-1a and sentinel-2 imagery based on the cnn model. Agronomy 13(7) (2023) https://doi.org/10.3390/agronomy13071723

    2023 doi

  9. [16]

    arXiv preprint arXiv:2409.09451 (2024)

    Chang, Y.-C., Stewart, A.J., Bastani, F., Wolters, P., Kannan, S., Huber, G.R., Wang, J., Banerjee, A.: On the generalizability of foundation models for crop type mapping. arXiv preprint arXiv:2409.09451 (2024)

    2024 arXiv

  10. [17]

    Remote Sensing of Environment 264, 112603 (2021)

    Turkoglu, M.O., D’Aronco, S., Perich, G., Liebisch, F., Streit, C., Schindler, K., Wegner, J.D.: Crop mapping from image time series: Deep learning with multi- scale label hierarchies. Remote Sensing of Environment 264, 112603 (2021)

    2021

  11. [18]

    Scientific Data 12(1), 664 (2025)

    Reuss, J., Macdonald, J., Becker, S., Richter, L., K¨ orner, M.: The eurocropsml time series benchmark dataset for few-shot crop type classification in europe. Scientific Data 12(1), 664 (2025)

    2025

  12. [19]

    arXiv preprint arXiv:1905.11893 (2019)

    Rußwurm, M., Pelletier, C., Zollner, M., Lef` evre, S., K¨ orner, M.: Breizhcrops: A time series dataset for crop type mapping. arXiv preprint arXiv:1905.11893 (2019)

    2019 arXiv

  13. [20]

    Remote sensing of Environment 202, 18–27 (2017)

    Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., Moore, R.: Google earth engine: Planetary-scale geospatial analysis for everyone. Remote sensing of Environment 202, 18–27 (2017)

    2017

  14. [21]

    Remote Sensing of Environment 210, 35–47 (2018) https://doi.org/10.1016/j.rse.2018.02.045

    Cai, Y., Guan, K., Peng, J., Wang, S., Seifert, C., Wardlow, B., Li, Z.: A high- performance and in-season classification system of field-level crop types using time-series landsat data and a machine learning approach. Remote Sensing of Environment 210, 35–47 (2018) https://do...

    2018 doi

  15. [22]

    Remote Sensing 7, 16091–16107 (2015) https://doi

    Li, Q., Wang, C., Lu, L.: Object-based crop classification with landsat-modis enhanced time-series data. Remote Sensing 7, 16091–16107 (2015) https://doi. org/10.3390/rs71215820

    2015 doi

  16. [23]

    Journal of Agricultural Meteorology 72, 1–11 (2016) https://doi

    Tatsumi, K., Yamashiki, Y., Morante, A., Ramos-Fern´ andez, L., Apaclla, R.: Pixel-based crop classification in peru from landsat 7 etm+ images using a random forest model. Journal of Agricultural Meteorology 72, 1–11 (2016) https://doi. 21 org/10.2480/agrmet.D-15-00010

    2016 doi

  17. [24]

    In: IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium, pp

    Song, J., Xing, M., Ma, Y., Wang, L., Luo, K., Quan, X.: Crop classification using multitemporal landsat 8 images. In: IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium, pp. 2407–2410 (2019). https://doi. org/10.1109/IGARSS.2019.8899274

    2019

  18. [25]

    In: 2012 IEEE International Geoscience and Remote Sensing Symposium, pp

    Spoto, F., Sy, O., Laberinti, P., Martimort, P., Fernandez, V., Colin, O., Hoersch, B., Meygret, A.: Overview of sentinel-2. In: 2012 IEEE International Geoscience and Remote Sensing Symposium, pp. 1707–1710 (2012). IEEE

    2012

  19. [26]

    Remote sensing of environment 120, 9–24 (2012)

    Torres, R., Snoeij, P., Geudtner, D., Bibby, D., Davidson, M., Attema, E., Potin, P., Rommen, B., Floury, N., Brown, M., et al.: Gmes sentinel-1 mission. Remote sensing of environment 120, 9–24 (2012)

    2012

  20. [27]

    Remote Sensing of Environment 269, 112831 (2022) https://doi.org/10.1016/j.rse.2021.112831

    Blickensd¨ orfer, L., Schwieder, M., Pflugmacher, D., Nendel, C., Erasmi, S., Hostert, P.: Mapping of crop types and crop sequences with combined time series of sentinel-1, sentinel-2 and landsat 8 data for germany. Remote Sensing of Environment 269, 112831 (2022) https://doi....

    2022

  21. [28]

    International Journal of Remote Sensing 40(17), 6553–6595 (2019) https://doi.org/10.1080/01431161

    Orynbaikyzy, A., Gessner, U., and, C.C.: Crop type classification using a combi- nation of optical and radar remote sensing data: a review. International Journal of Remote Sensing 40(17), 6553–6595 (2019) https://doi.org/10.1080/01431161. 2019.1569791 https://doi.org/10.1080/0...

    2019

  22. [29]

    In: Thirty-fifth Conference on Neural Infor- mation Processing Systems Datasets and Benchmarks Track (Round 2) (2021)

    Tseng, G., Zvonkov, I., Nakalembe, C.L., Kerner, H.: Cropharvest: A global dataset for crop-type classification. In: Thirty-fifth Conference on Neural Infor- mation Processing Systems Datasets and Benchmarks Track (Round 2) (2021)

    2021

  23. [30]

    : Fields of the world: A machine learning benchmark dataset for global agricultural field boundary segmentation

    Kerner, H., Chaudhari, S., Ghosh, A., Robinson, C., Ahmad, A., Choi, E., Jacobs, N., Holmes, C., Mohr, M., Dodhia, R., et al. : Fields of the world: A machine learning benchmark dataset for global agricultural field boundary segmentation. In: Proceedings of the AAAI Conference...

    2025

  24. [31]

    Geocarto International 26(5), 341–358 (2011) https://doi.org/10

    Boryan, C., Yang, Z., Mueller, R., and, M.C.: Monitoring us agriculture: the us department of agriculture, national agricultural statistics service, cropland data layer program. Geocarto International 26(5), 341–358 (2011) https://doi.org/10. 1080/10106049.2011.562309 https://...

    2011

  25. [33]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Tseng, G., Kerner, H., Nakalembe, C., Becker-Reshef, I.: Learning to predict crop type from heterogeneous sparse labels using meta-learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 22 1111–1120 (2021)

    2021

  26. [34]

    arXiv preprint arXiv:2311.15138 (2023)

    Gurav, R., Patel, H., Shang, Z., Eldawy, A., Chen, J., Scudiero, E., Papalexakis, E.: Can sam recognize crops? quantifying the zero-shot performance of a seman- tic segmentation foundation model on generating crop-type maps using satellite imagery for precision agriculture. ar...

    2023 arXiv

  27. [35]

    Remote Sensing 14(10), 2404 (2022)

    Mahlayeye, M., Darvishzadeh, R., Nelson, A.: Cropping patterns of annual crops: A remote sensing review. Remote Sensing 14(10), 2404 (2022)

    2022

  28. [36]

    International Journal of Remote Sensing 25 (2004) https://doi.org/10.1080/01431160310001619580

    Wit, A., Clevers, J.G.P.W.: Efficiency and accuracy of per-field classification for operational crop mapping. International Journal of Remote Sensing 25 (2004) https://doi.org/10.1080/01431160310001619580

    2004 doi

  29. [37]

    Remote Sensing 16(24) (2024) https://doi.org/10.3390/rs16244620

    Vanpoucke, K., Heremans, S., Buls, E., Somers, B.: Field-level classification of winter catch crops using sentinel-2 time series: Model comparison and transfer- ability. Remote Sensing 16(24) (2024) https://doi.org/10.3390/rs16244620

    2024 doi

  30. [38]

    Remote Sensing 16(1), 127 (2023)

    Abdali, E., Valadan Zoej, M.J., Taheri Dehkordi, A., Ghaderpour, E.: A parallel- cascaded ensemble of machine learning models for crop type classification in google earth engine using multi-temporal sentinel-1/2 and landsat-8/9 remote sensing data. Remote Sensing 16(1), 127 (2023)

    2023

  31. [39]

    Remote sensing of environment 210, 12–24 (2018)

    Yin, H., Prishchepov, A.V., Kuemmerle, T., Bleyhl, B., Buchner, J., Radeloff, V.C.: Mapping agricultural land abandonment from spatial and temporal seg- mentation of landsat time series. Remote sensing of environment 210, 12–24 (2018)

    2018

  32. [40]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, 474–487 (2021) https://doi.org/10.1109/ JSTARS.2020.3036602

    Yuan, Y., Lin, L.: Self-supervised pretraining of transformers for satellite image time series classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, 474–487 (2021) https://doi.org/10.1109/ JSTARS.2020.3036602

    2021

  33. [41]

    https://arxiv.org/abs/2411.05359

    Dua, R., Saxena, N., Agarwal, A., Wilson, A., Singh, G., Tran, H., Deshpande, I., Kaur, A., Aggarwal, G., Nath, C., Basu, A., Batchu, V., Holla, S., Kurle, B., Missura, O., Aggarwal, R., Garg, S., Shah, N., Singh, A., Tewari, D., Dondzik, A., Adsul, B., Sohoni, M., Praveen, A....

    2024 arXiv

  34. [42]

    http://s2geometry.io/resources/s2cell statistics.html

    S2 Geometry. http://s2geometry.io/resources/s2cell statistics.html

  35. [43]

    https://arxiv.org/ abs/1706.03762 23

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention Is All You Need (2023). https://arxiv.org/ abs/1706.03762 23

    2023 arXiv

  36. [44]

    https://arxiv.org/abs/2206.06488

    Xu, P., Zhu, X., Clifton, D.A.: Multimodal Learning with Transformers: A Survey (2023). https://arxiv.org/abs/2206.06488

    2023 arXiv

  37. [45]

    https://arxiv.org/abs/1901.07017

    Watters, N., Matthey, L., Burgess, C.P., Lerchner, A.: Spatial Broadcast Decoder: A Simple Architecture for Learning Disentangled Representations in V AEs (2019). https://arxiv.org/abs/1901.07017

    2019 arXiv

  38. [46]

    https://arxiv.org/abs/2301.08871

    Li, Z., Rao, Z., Pan, L., Wang, P., Xu, Z.: Ti-MAE: Self-Supervised Masked Time Series Autoencoders (2023). https://arxiv.org/abs/2301.08871

    2023 arXiv

  39. [47]

    https://arxiv.org/abs/2111.06377

    He, K., Chen, X., Xie, S., Li, Y., Doll´ ar, P., Girshick, R.: Masked Autoencoders Are Scalable Vision Learners (2021). https://arxiv.org/abs/2111.06377

    2021 arXiv

  40. [48]

    Journal of Applied Remote Sensing 15, 024513–024513 (2021)

    Soler-P´ erez-Salazar, M.J., Ortega-Garc ´ ıa, N., Vaca-Mier, M., Cram-Hyedric, S.: Maize and sorghum field segregation using multi-temporal sentinel-2 data in central mexico. Journal of Applied Remote Sensing 15, 024513–024513 (2021)

    2021

  41. [49]

    https://upag.gov.in/

    Unified Portal For Agricultural Statistics. https://upag.gov.in/

  42. [50]

    Agronomy Journal 115(3), 1214–1236 (2023)

    Worrall, G., Judge, J., Boote, K., Rangarajan, A.: In-season crop phenology using remote sensing and model-guided machine learning. Agronomy Journal 115(3), 1214–1236 (2023)

    2023

  43. [51]

    https: //arxiv.org/abs/1412.6980

    Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization (2017). https: //arxiv.org/abs/1412.6980

    2017 arXiv

  44. [52]

    Lin, T.-Y., Goyal, P., Girshick, R., He, K., Doll´ ar, P.: Focal Loss for Dense Object Detection (2018). https://arxiv.org/abs/1708.02002 24 Appendix A Ground Truth Data A.1 Data Collection Process The ground truth labels licensed from partners were collected by following oper...

    2018 arXiv

  45. [53]

    Surveyors physically visited fields where there was an active crop

  46. [54]

    The GPS location of the surveyors was recorded from mobile devices they were required to carry

  47. [55]

    Surveyors recorded the label of the crop by observing the crop themselves

  48. [56]

    A.2 Crop Label Distribution Distribution of the labels is visualized in Fig

    The time of collection of label was automatically registered. A.2 Crop Label Distribution Distribution of the labels is visualized in Fig. A1. Wheat Sugarcane Soybeans Mustard Corn Rice Cotton Gram Sorghum Groundnut Chilli Bajra Others 0 2000 4000 6000 8000 10000 12000 14000Nu...

    2000

  49. [57]

    We take the median per band across all locations in the field, similar to Sec

    We start with Sentinel-2 time-series for the field. We take the median per band across all locations in the field, similar to Sec. 3.2

  50. [58]

    This is again a time- series

    We compute the NDVI using the formula suggested in [10]. This is again a time- series

  51. [59]

    We drop all NDVI values where the cs cdf band from CloudScore+ is < 0.6, as suggested in the CloudScore+ manual

    CloudScore+ from Earth Engine ([20]) is used to identify where the NDVI is corrupted by clouds. We drop all NDVI values where the cs cdf band from CloudScore+ is < 0.6, as suggested in the CloudScore+ manual

  52. [60]

    A moving average is applied to the NDVI signal with a window of 20 days

  53. [61]

    The smoothened NDVI is thresholded to get a time-series marking each time step in the time-series as bare or vegetated, where we declare the field to be vegetated at time step t if the corresponding NDVI is >= 0.4 and bare otherwise

  54. [62]

    This binary time series is then segmented to find continuous vegetated periods

  55. [63]

    broadcasted

    We further apply post-processing to merge two identified seasons if the gap between them is less than 1 month, as this may arise due to signal corruption and is unlikely to happen in the real world. Thus for each field, we get a series of crop seasons, each defined as ( tstart...

    2023

  56. [64]

    We select the administrative boundary of interest (in our case, we use state boundaries)

  57. [65]

    We use an ensemble of the three models evaluated in the 4

    We run our crop classification pipeline (season detection as well as crop classifica- tion) for all fields from ALU within this boundary. We use an ensemble of the three models evaluated in the 4. The prediction probabilities per crop are averaged over the models

  58. [66]

    When the pipeline is run for date t, the models are given access to satellite signals from before date t only

    The pipeline is run independently for the 1 st of each month for the desired time period (Jan 2022 - Dec 2024 in our case). When the pipeline is run for date t, the models are given access to satellite signals from before date t only. 30

    2022

  59. [67]

    Each prediction contains the season estimate

    The predictions for each field are then combined based on the identified crop sea- sons. Each prediction contains the season estimate. We group all overlapping season estimates and take the latest prediction from this group

  60. [68]

    The predictions for each field are marked with the season (Winter or Monsoon) by the estimated start date of the season (May-October is Monsoon, November-March is Winter)

  61. [69]

    NaN” values for precision indicate that no predictions were made. “NaN

    The total area under each crop is then summed up over all fields, grouped by the predicted crop. 31 (a) Winter (Rabi) T otal Wheat Gram Mustard Others 0 2000 4000 6000 8000 10000 12000Area (Thousand Hectares) MadhyaPradesh Census Predicted T otal Wheat Others Mustard Gram 0 20...

    2000

Pith tools

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