Pith. sign in

REVIEW 4 major objections 7 minor 26 references

STS-NET: Spatio-Temporal Stress Network for Self-Supervised Crop Stress Detection using Satellite Image Time Series

T0 review · 4 major / 7 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read STS-NET shows that a self-supervised 3D autoencoder trained on unlabeled satellite time series can detect sugarcane water stress with 97.98% precision.

desk verdict The crop-stress direction is reasonable, but the pixel-level split on 62 tiny plots makes the headline accuracies untrustworthy. read the letter →

arxiv 2607.18791 v1 pith:ALVZIFQ7 submitted 2026-07-21 cs.CV

Pradeep Dalal , Rajiv Ranjan , Sushil Ghildiyal , Shashank Tamaskar , Neeraj Goel This is my paper
classification cs.CV
keywords cropstressdetectionself-supervisedlearning3Dconvolutionalautoencodersatelliteimagetimeseriesvegetationindicessugarcanespatio-temporalfeaturesPlanetScopeimagery
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 tries to show that a crop-stress detector can be built almost without labeled data. The authors pretrain a 3D convolutional autoencoder to reconstruct pixel-level satellite time series from unlabeled imagery, then freeze the encoder and train simple classifiers on the resulting features. On a controlled sugarcane farm with 62 plots, the approach reports 97.98% precision for water stress, 85.08% for nitrogen stress, and 83.47% for combined water-nitrogen stress. The point that matters: if the results hold, large unlabeled satellite archives can substitute for expensive field labels in monitoring crop health.

What carries the argument

The central object is the 3D convolutional autoencoder operating on pixel-level satellite image time series. Input is a tensor of shape 1×1×T×C: one pixel tracked across T=32 timesteps with C=5 channels (red, green, blue, near-infrared, and the mean of NDVI, GNDVI, RECI, and NDRE). The encoder compresses this to a 64-dimensional latent vector, and pretraining minimizes MSE between input and reconstructed series. In the fine-tuning stage the encoder is frozen and the latent vector is treated as a feature extractor for Random Forest and XGBoost classifiers, with water, nitrogen, and combined stress as separate multiclass tasks. The mechanism that carries the argument is the claim that reconstr

What would settle it

Re-run the evaluation with a plot-level split, holding out entire water-nitrogen treatment plots (e.g., leave-one-plot-out across the 62 fields) and retrain the lightweight classifiers on the same frozen encoder features. If water-stress precision falls well below 97.98%, the current numbers reflect within-plot memorization. A second check: evaluate on a different sugarcane farm in a different season without retraining.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that self-supervised reconstruction pretraining turns a small 3D-CAE into a feature extractor for crop stress. The encoder is trained to reconstruct pixel-level 32-step time series (RGB, NIR, and the mean of four vegetation indices) from an unlabeled region in Punjab, then frozen. Labeled pixels from 62 sugarcane plots in Uttar Pradesh are encoded into 64-dimensional vectors that feed Random Forest or XGBoost. The authors report this beats supervised baselines and a pretrained transformer baseline with only 19.2K parameters, and stays accurate with just 8 time steps. They conclude that spatio-temporal stress patterns transfer from unlabeled

Load-bearing premise

The results rest on the assumption that an 80:20 random split of pixel-level time series from 62 small, spatially contiguous plots produces independent training and test samples; because pixels from the same plot are correlated in space and time, the reported accuracies may overstate how well the model generalizes to a new field.

Editorial extensions

If this is right

  • A stress detector can be built with minimal labeled data: the 19.2K-parameter encoder is pretrained without labels and only the lightweight head needs supervision.
  • The method degrades gracefully with shorter time series: performance with 8 timesteps (about two months) remains around 90% water-stress accuracy, suggesting applicability to shorter-duration crops.
  • The same frozen features support three separate tasks—water, nitrogen, and combined nine-class water-nitrogen stress—so one pretrained encoder can serve multiple monitoring goals.
  • Because the encoder is a generic feature extractor, it can be paired with inexpensive classifiers such as Random Forest and XGBoost rather than requiring a large deep network.
  • Reported accuracies are higher than supervised baselines and a pretrained transformer baseline, implying that domain-specific self-supervised pretraining on agricultural time series is a key ingredient.

Reading between the lines

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

  • The pixel-level 80:20 split is taken from a single 2.5-acre farm with 62 adjacent, uniformly treated plots. A fairer test would withhold whole plots; if accuracies drop, part of the reported performance reflects spatial autocorrelation rather than transferable stress detection.
  • Since pretraining uses only unlabeled satellite imagery, the same encoder could be evaluated on other crops and regions without new labels; the paper's own 8-timestep result suggests a quick transfer test on shorter-season crops.
  • A natural deployment reading, not tested here: because the inputs are per-pixel time series, the model could in principle produce weekly stress maps for an entire field, turning detection into a monitoring tool rather than a one-time classification.
  • The mean-of-four-vegetation-indices channel rather than individual indices may be a deliberate compression; a testable extension is whether keeping the indices separate improves interpretability of which stress dominates.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes STS-NET, a self-supervised framework for crop stress detection from PlanetScope satellite image time series (SITS). A 3D convolutional autoencoder (3D-CAE) is pretrained on a new unlabeled dataset (BSPT) from Barnala, Punjab, using a pixel-reconstruction objective. The encoder is then frozen and used to extract 64-dimensional features from labeled pixels in the Lakhimpur-Kheri (LK) sugarcane experiment, which contains 62 plots with nine water-by-nitrogen treatment combinations. These features are classified by Random Forest, XGBoost, or SVM for water stress (3-class), nitrogen stress (3-class), and combined stress (9-class). The paper reports strong results, e.g., water-stress precision of 97.98% (Table I), and claims that self-supervised pretraining on unlabeled SITS yields a robust, lightweight feature extractor with minimal reliance on labeled data.

Significance. If the empirical claims were sound, the paper would make a useful contribution: a lightweight (19.2K parameters) self-supervised encoder for crop-stress detection, a new unlabeled SITS dataset, and a comparison against a strong self-supervised baseline (Presto). However, the central empirical claim is not supported as stated. The evaluation uses a pixel-level random split of a small, spatially autocorrelated experimental farm, which leaks label information into the test set; no ablation isolates the effect of self-supervised pretraining; and the model as described in Eq. (1) operates on a single pixel, so the claimed spatial component is degenerate. These are not presentation issues but load-bearing problems with the evidence and with the method's stated novelty.

major comments (4)
  1. [Section V-A/V-B] The 80:20 train/test split is performed at the pixel level on a dataset of 62 small, contiguous, uniformly treated plots (approximately 0.05 acres each). With PlanetScope's ~3 m resolution, each plot is only a few pixels, and within-plot spatial autocorrelation is strong. Randomly assigning pixels to train/test places pixels from the same treatment plot in both sets, allowing the classifier to memorize plot-specific radiometric signatures rather than learn generalizable stress patterns. The near-perfect water-stress precision (Table I, 0.98) is consistent with this leakage. A plot-level split, such as leave-one-plot-out, is required. This issue invalidates the reported accuracies, precisions, and all cross-model comparisons.
  2. [Section V-B, Table I] There is no ablation that trains the same 3D-CAE encoder from scratch on the LK labeled data (or with random initialization) before attaching the classifier. The supervised CNN, SVM, RF, and XGBoost baselines use raw input features, and Presto is a different architecture. Consequently, the paper does not isolate whether the reported gains come from self-supervised pretraining on BSPT, from the 3D-CAE architecture itself, or from the specific input representation. A control with the same encoder but without pretraining is essential to support the claim that self-supervised pretraining is the source of the improvement.
  3. [Eq. (1), Fig. 2] The input tensor is defined as x ∈ R^{1×1×T×C}, i.e., a single pixel tracked over T=32 timesteps with C=5 channels. With spatial dimensions of 1×1, the 3D convolutions do not integrate any spatial neighborhood. The paper's repeated claims of learning 'spatio-temporal' stress patterns, and the architecture name STS-NET, are therefore not supported by the described model. If spatial context is intended, the input must be image patches (e.g., H×W×T×C); if not, the method should be described as temporal-only and the claims revised accordingly.
  4. [Section V-B, Table II] The optimal temporal depth T and input channel configuration are selected using ablations performed on the same small LK dataset whose test split is already contaminated by the pixel-level split described above. This selection-on-test further biases the reported numbers. The paper should use a separate validation split or nested cross-validation for hyperparameter and configuration selection, and report metrics on a truly held-out set of fields or farms.
minor comments (7)
  1. [Index Terms] Typos: 'Saptio-temporal' should be 'Spatio-temporal' and 'T erms' should be 'Terms'.
  2. [Abstract vs Section III-A] The abstract states a '2.5-acre test plot', while Section III-A and Fig. 1 state approximately 3 acres. Please correct the inconsistency.
  3. [Table I] Several entries in the proposed-method rows are formatted without spacing (e.g., '0.970.98', '0.960.95'), making the table hard to read. Please clean up the formatting.
  4. [Eq. (1)] The notation for t_i is ambiguous. Please clarify that each t_i is a 5-dimensional vector containing R, G, B, NIR, and the average of the four vegetation indices, and state whether the average is computed over the four indices at the same timestamp.
  5. [Section IV-A vs V-A] Section IV-A mentions 'stratified sampling applied during train-validation splitting', while Section V-A describes an 80:20 training/testing split. Please clarify how validation is used, especially with early stopping, and whether the same split is used for all classifiers.
  6. [General] No statement is provided on data or code availability. For reproducibility, please indicate whether the BSPT dataset and implementation will be released.
  7. [References] Reference [22] (Presto) is cited as an arXiv preprint. If a peer-reviewed version exists, please cite the published version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the self-supervised pretraining and supervised evaluation chain is independent.

full rationale

The claimed derivation chain is not circular. STS-NET pretrains a 3D-CAE on the unlabeled BSPT dataset with an MSE reconstruction objective—this objective does not use stress labels or treatments. The fine-tuning stage attaches a classifier to frozen encoder features and trains it on the LK dataset's experimentally defined labels (water/nitrogen treatments), which is standard supervised evaluation. No equation in the paper defines stress labels from the input vegetation indices; labels originate from controlled experimental conditions. Self-citations (e.g., [23]-[26]) are used to describe the LK dataset and prior 3D-CAE work, but they are not invoked as proof of the method's effectiveness. The reported accuracies are empirical results, not quantities forced by construction. Concerns about pixel-level splitting and ablation-driven configuration selection are validity/generalization issues rather than circularity: they do not make the predicted stress labels equivalent to the model inputs or to the pretraining objective. Therefore no specific circular step can be exhibited.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on untested transfer assumptions (Barnala pretraining to LK), uniform pixel-level label propagation from field treatments, an independence assumption for pixel sampling, and the assumption that reconstruction pretraining produces stress-relevant features. The BSPT dataset is new but is used only for pretraining and is not independently validated.

free parameters (7)
  • Input channel configuration (RGB+NIR+Avg(VIs)) = 5 channels: R, G, B, NIR, mean of NDVI, GNDVI, RECI, NDRE
    Selected based on ablation on the LK dataset (Table II), not on an independent validation set.
  • Temporal depth T = 32 time steps
    Ablation Table II shows T=32 gives the best accuracy on the LK benchmark; tuned to the test set.
  • Learning rate = 0.001
    Chosen via 'extensive experimentation' (Section V-A); no grid search or sensitivity results shown.
  • Batch size = 256
    Chosen via 'extensive experimentation' (Section V-A).
  • Latent dimension = 64
    The encoder outputs a 64-dimensional latent vector; no justification or ablation.
  • Training epochs / early stopping = up to 40 epochs with early stopping
    Training schedule; early stopping criterion not specified.
  • Vegetation index averaging = average of NDVI, GNDVI, RECI, NDRE
    t_VIs = (1/4) * sum of four indices; selected because the ablation showed best performance (Table II).
assumptions (5)
  • domain assumption Field-level water and nitrogen treatments are uniform within each of the 62 LK plots, so every pixel in a plot can be labeled with the plot's treatment.
    Section III-A describes treatments assigned per block/plot; pixel-level label propagation is assumed without verification.
  • domain assumption Pretraining on 196 km² of Barnala, Punjab agricultural land transfers to the LK site in Uttar Pradesh.
    The pretraining and evaluation regions differ in crop mix, climate, and soil; no domain-shift analysis is provided.
  • domain assumption Pixel-level 80:20 stratified split yields independent training and test samples.
    Section V-A describes a split of pixel-level time-series vectors; spatial autocorrelation within small, contiguous, uniformly treated plots is not addressed.
  • domain assumption Mean imputation of missing pixel values preserves stress-related temporal signal.
    Section IV-A says missing values were imputed with a mean-based strategy; no sensitivity analysis is given.
  • domain assumption MSE reconstruction on unlabeled SITS learns features relevant to downstream stress classification.
    This is the standard self-supervised transfer assumption, but the paper never tests it by comparing a pretrained 3D-CAE to a randomly initialized 3D-CAE.

how reviews work

0 comments
Cite this review

Pith. "Pith review of STS-NET: Spatio-Temporal Stress Network for Self-Supervised Crop Stress Detection using Satellite Image Time Series." pith.science (2026). https://pith.science/paper/ALVZIFQ7

@misc{pith2026260718791,
  author       = {Pith},
  title        = {Pith review of: STS-NET: Spatio-Temporal Stress Network for Self-Supervised Crop Stress Detection using Satellite Image Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ALVZIFQ7}},
  note         = {Machine review of arXiv:2607.18791}
}
read the original abstract

Early and accurate detection of crop stress is essential to improve agricultural productivity and ensure global food security. However, collecting a large labeled crop stress dataset is a challenging task. To address this challenge, we introduce a novel spatial-temporal stress network (STS-NET), built on a self-supervised 3D-convolutional autoencoder (3D-CAE), designed to utilize Satellite Image Time Series (SITS) data for crop stress detection. STS-NET exploits four vegetation indices: Normalized Difference Vegetation Index (NDVI), Normalized Difference Vegetation Index (GNDVI), Red-Edge Chlorophyll Index (RECI) and Normalized Difference Red-Edge Index (NDRE) obtained from high resolution Planetscope imagery to capture spatiotemporal stress patterns. The model is trained on our BSPT (Barnala Spatial-Temporal) dataset and evaluated on a real-world sugarcane dataset collected over a year from a 2.5-acre test plot located in Lakhimpur-Kheri (LK) district in Uttar Pradesh in India. STS-NET achieved a precision of 97. 98\% for water stress, 85.08\% for nitrogen stress, and 83.47\% for combined stress. The results demonstrate the potential of STS-NET in effectively detecting stress in sugarcane crops with minimal reliance on labeled data. Furthermore, STS-NET can serve as a robust feature extractor for simpler models.

Figures

Figures reproduced from arXiv: 2607.18791 by the authors.

Figure 1
Figure 1. (a) BSPT dataset (Punjab, India) collected from PlanetScope imagery, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Represents the STS-NET architecture pipeline for stress detection in sugarcane farms. In pretraining stage, the autoencoder was trained on unlabelled [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 1 linked inside Pith

  1. [1]

    Performance of vegetation indices to estimate green biomass accumulation in common bean,

    T. O. C. Barboza, M. Ardigueri, G. F. C. Souza, M. A. J. Ferraz, J. R. F. Gaudencio, and A. F. d. Santos, “Performance of vegetation indices to estimate green biomass accumulation in common bean,”AgriEngineering, vol. 5, no. 2, pp. 840–854, 2023

    2023

  2. [2]

    Determining nitrogen deficiencies for maize using various remote sensing indices,

    B. W. Burns, V. S. Green, A. A. Hashem, J. H. Massey, A. M. Shew, M. A. A. Adviento-Borbe, and M. Milad, “Determining nitrogen deficiencies for maize using various remote sensing indices,”Precision Agriculture, vol. 23, no. 3, pp. 791–811, 2022

    2022

  3. [3]

    Application of ndvi and ndre vege- tation indices in the assessment of soybean productivity under nitrogen controlled-release fertilizer,

    B. Boiarskii and M. Sinegovskii, “Application of ndvi and ndre vege- tation indices in the assessment of soybean productivity under nitrogen controlled-release fertilizer,” in2022 VIII International Conference on Information Technology and Nanotechnology (ITNT). IEEE, 2022, pp. 1–6

    2022

  4. [4]

    Stressnet: a spatial-spectral-temporal deformable attention-based framework for water stress classification in maize,

    T. Nampally, K. Kumar, S. Chatterjee, R. Pachamuthu, B. Naik, and U. B. Desai, “Stressnet: a spatial-spectral-temporal deformable attention-based framework for water stress classification in maize,”Frontiers in Plant Science, vol. 14, p. 1241921, 2023

    2023

  5. [5]

    Mapping maize water stress based on uav multispectral remote sensing,

    L. Zhang, H. Zhang, Y. Niu, and W. Han, “Mapping maize water stress based on uav multispectral remote sensing,”Remote Sensing, vol. 11, no. 6, p. 605, 2019

    2019

  6. [6]

    Use of sugarcane industrial by-products for improving sugarcane productivity and soil health,

    M. Dotaniya, S. Datta, D. Biswas, C. Dotaniya, B. Meena, S. Rajendiran, K. Regar, and M. Lata, “Use of sugarcane industrial by-products for improving sugarcane productivity and soil health,”International Journal of Recycling of Organic Waste in Agriculture, vol. 5, pp. 185–194, 2016

    2016

  7. [7]

    Factors affecting the production of sugarcane yield and sucrose accumulation: suggested potential biological solutions,

    F. Mehdi, Z. Cao, S. Zhang, Y. Gan, W. Cai, L. Peng, Y. Wu, W. Wang, and B. Yang, “Factors affecting the production of sugarcane yield and sucrose accumulation: suggested potential biological solutions,”Frontiers in Plant Science, vol. 15, p. 1374228, 2024

    2024

  8. [8]

    Prediction of sugarcane sucrose content and optimal harvest date using multi-spectral time series image processing of satellite data,

    R. Ranjan and S. Tamaskar, “Prediction of sugarcane sucrose content and optimal harvest date using multi-spectral time series image processing of satellite data,” inIGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, 2024, pp. 4239–4244

    2024

Show all 26 references
  1. [9]

    Remote sensing crop water stress determination using cnn-vit architecture,

    K. Lehouel, C. Saber, M. Bouziani, and R. Yaagoubi, “Remote sensing crop water stress determination using cnn-vit architecture,”AI, vol. 5, no. 2, pp. 618–634, 2024

    2024

  2. [10]

    Estimation of maize properties and differentiating moisture and nitrogen deficiency stress via ground–based remotely sensed data,

    A. H. Elmetwalli and A. N. Tyler, “Estimation of maize properties and differentiating moisture and nitrogen deficiency stress via ground–based remotely sensed data,”Agricultural Water Management, vol. 242, p. 106413, 2020

    2020

  3. [11]

    Improved water management practices improve cropping system profitability and smallholder farmers’ incomes,

    S. Dutta, A. M. Laing, S. Kumar, M. K. Gathala, A. K. Singh, D. Gaydon, and P. Poulton, “Improved water management practices improve cropping system profitability and smallholder farmers’ incomes,”Agricultural Water Management, vol. 242, p. 106411, 2020

    2020

  4. [12]

    Crop water stress detection using remote sensing techniques,

    M. Safdar, M. A. Shahid, A. Sarwar, F. Rasul, M. D. Majeed, and R. M. Sabir, “Crop water stress detection using remote sensing techniques,” Environmental Sciences Proceedings, vol. 25, no. 1, p. 20, 2023

    2023

  5. [13]

    Imaging spectroscopy for estimating sugarcane leaf nitrogen concentration,

    E. M. Abdel-Rahman, F. B. Ahmed, and M. Van den Berg, “Imaging spectroscopy for estimating sugarcane leaf nitrogen concentration,” in Remote Sensing for Agriculture, Ecosystems, and Hydrology X, vol. 7104. SPIE, 2008, pp. 263–274

    2008

  6. [14]

    Crop water stress index computation approaches and their sensitivity to soil water dynamics,

    A. Katimbo, D. R. Rudnick, K. C. DeJonge, T. H. Lo, X. Qiao, T. E. Franz, H. N. Nakabuye, and J. Duan, “Crop water stress index computation approaches and their sensitivity to soil water dynamics,”Agricultural Water Management, vol. 266, p. 107575, 2022

    2022

  7. [15]

    Critical soil moisture thresholds of plant water stress in terrestrial ecosystems,

    Z. Fu, P. Ciais, A. F. Feldman, P. Gentine, D. Makowski, I. C. Prentice, P. C. Stoy, A. Bastos, and J.-P. Wigneron, “Critical soil moisture thresholds of plant water stress in terrestrial ecosystems,”Science Advances, vol. 8, no. 44, p. eabq7827, 2022

    2022

  8. [16]

    Efficient vegetation indices for phenotyping of abiotic stress tolerance in tea plant (camellia sinensis (l.) kuntze),

    L. Samarina, L. Malyukova, N. Koninskaya, V. Malyarovskaya, A. Ryndin, W. Tong, E. Xia, and E. Khlestkina, “Efficient vegetation indices for phenotyping of abiotic stress tolerance in tea plant (camellia sinensis (l.) kuntze),”Heliyon, vol. 10, no. 15, 2024

    2024

  9. [17]

    Relationship between spectral vegetation indices and crop parameters of maize under nitrogen and water stress,

    K. Ramachandiran and S. Pazhanivelan, “Relationship between spectral vegetation indices and crop parameters of maize under nitrogen and water stress,”Journal of Agrometeorology, vol. 19, no. 2, pp. 114–119, 2017

    2017

  10. [18]

    Wheat crop growth monitoring using multi-spectral vegetation indices in bhal region, gujarat state,

    M. Parmar, H. Solanki, and M. H. Kalubarme, “Wheat crop growth monitoring using multi-spectral vegetation indices in bhal region, gujarat state,”International Journal of Advances in Agricultural Science and Technology, vol. 7, no. 9, pp. 47–66, 2020

    2020

  11. [19]

    Field-scale estimation of sugarcane leaf nitrogen content using vegetation indices and spectral bands of sentinel-2: Application of random forest and support vector regression,

    M. Soltanikazemi, S. Minaei, H. Shafizadeh-Moghadam, and A. Mah- davian, “Field-scale estimation of sugarcane leaf nitrogen content using vegetation indices and spectral bands of sentinel-2: Application of random forest and support vector regression,”Computers and Electronics ...

    2022

  12. [20]

    Heavy metal-induced stress in rice crops detected using multi-temporal sentinel-2 satellite images,

    M. Liu, T. Wang, A. K. Skidmore, and X. Liu, “Heavy metal-induced stress in rice crops detected using multi-temporal sentinel-2 satellite images,” Science of the total environment, vol. 637, pp. 18–29, 2018

    2018

  13. [21]

    Denseresunet: An architecture to assess water-stressed sugarcane crops from sentinel-2 satellite imagery

    S. S. Virnodkar, V. K. Pachghare, V. C. Patil, and S. K. Jha, “Denseresunet: An architecture to assess water-stressed sugarcane crops from sentinel-2 satellite imagery.”Traitement du Signal, vol. 38, no. 4, 2021

    2021

  14. [22]

    Lightweight, pre-trained transformers for remote sensing timeseries,

    G. Tseng, R. Cartuyvels, I. Zvonkov, M. Purohit, D. Rolnick, and H. Kerner, “Lightweight, pre-trained transformers for remote sensing timeseries,” arXiv preprint arXiv:2304.14065, 2023

    2023 arXiv

  15. [23]

    3d-ctm: Unsupervised crop type mapping based on 3d convolutional autoencoder and satellite image time series,

    K. Singh, R. Ranjan, S. Ghildiyal, S. Tamaskar, and N. Goel, “3d-ctm: Unsupervised crop type mapping based on 3d convolutional autoencoder and satellite image time series,”IEEE Geoscience and Remote Sensing Letters, 2024

    2024

  16. [24]

    Evaluating sugarcane yield variability with uav-derived cane height under different water and nitrogen conditions,

    R. Ranjan, T. Birdh, N. Mandal, D. Kumar, and S. Tamaskar, “Evaluating sugarcane yield variability with uav-derived cane height under different water and nitrogen conditions,” inPattern Recognition. ICPR 2024 International Workshops and Challenges, S. Palaiahnakote, S. Schucke...

    2024

  17. [25]

    Yield forecasting of sugarcane using machine learning and uav-derived canopy height,

    R. Ranjan, A. Upadhyaya, D. Kumar, and S. Tamaskar, “Yield forecasting of sugarcane using machine learning and uav-derived canopy height,” in2024 IEEE India Geoscience and Remote Sensing Symposium (In- GARSS), 2024, pp. 1–4

    2024

  18. [26]

    Evaluation of computer vision pipeline for farm-level analytics: A case study in sugarcane,

    S. Shikhar, R. Ranjan, A. Sa, A. Srivastava, Y. Srivastava, D. Kumar, S. Tamaskar, and A. Sobti, “Evaluation of computer vision pipeline for farm-level analytics: A case study in sugarcane,” inProceedings of the 7th ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Soc...

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.