Pith. sign in

REVIEW 4 major objections 6 minor 20 references

Interpolation pour l'augmentation de donnees : Application \`a la gestion des adventices de la canne a sucre a la Reunion

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that interpolation—especially Gaussian processes with combined kernels—can serve as a cheap form of data augmentation for small geo-referenced ecological datasets.

desk verdict A sensible application of GP/kriging augmentation to a weed dataset, but the single random split makes the headline 'significantly improves' claim unproven. read the letter →

arxiv 2501.12400 v1 pith:M76XZOVO submitted 2025-01-10 q-bio.QM cs.LGstat.APstat.MEstat.ML

Frederick Fabre Ferber , Dominique Gay , Jean-Christophe Soulie , Jean Diatta , Odalric-Ambrym Maillard This is my paper · ORCID
classification q-bio.QMcs.LGstat.APstat.MEstat.ML
keywords dataaugmentationspatialinterpolationGaussianprocessregressionkrigingweedcoversugarcaneCommelinabenghalensisgeo-referenced
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 asks whether interpolation can serve as data augmentation when field data are expensive and geo-referenced. Using 745 sugarcane plots on La Réunion with the weed species Commelina benghalensis as the target, the authors generate synthetic observations by Gaussian-process regression and by kriging, then train seven regression algorithms on the augmented sets. The central claim is that GP-based interpolation, especially with a combined kernel (GP-COMB), consistently lowers prediction error compared with the unaugmented dataset while needing fewer added points than kriging variants. On the strongest cases, GP-COMB cuts mean-squared error from 36.45 to 13.27 for Gradient Boosting, from 23.83 to 13.35 for Random Forest, and from 16.30 to 13.17 for k-NN. A secondary finding is that kriging produces more spatially homogeneous weed-cover maps, which may be preferable when smooth spatial coverage matters more than raw predictive accuracy.

What carries the argument

The carrying mechanism is the covariance model used to predict target values at unobserved locations: Gaussian process regression with a kernel (linear, RBF, quadratic, or a BIC-selected combination built by sum and product operators) and kriging with a variogram (linear, exponential, Gaussian, spherical). New points are placed in zones likely to contain sugarcane, auxiliary covariates are retrieved from a meteorological service, and the interpolation method assigns each new point a weed-cover value from the spatial covariance structure. The added rows then enter the training set of downstream regressors (linear, ridge, SVR, random forest, gradient boosting, k-NN, MLP) whose validation MSE is measured with 200 extra points and, for the best model, with 0 to 300 extra points.

What would settle it

Re-run the same protocol with spatial blocking, for example excluding all test plots within a set distance (say 1 km) of any training plot, and compare GP-COMB against the unaugmented base dataset. If the MSE advantage disappears or reverses, the central claim fails; if it persists, the claim survives the leakage objection.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that interpolation with Gaussian processes is an effective and sample-efficient data-augmentation strategy for small geo-referenced ecological datasets. The combined-kernel GP (GP-COMB), built by searching over sums and products of linear, RBF, and quadratic kernels with a BIC criterion, achieved the lowest validation MSE for Random Forest, Gradient Boosting, and k-NN among all eight interpolation methods tested, and all methods improved on the baseline dataset. The paper further reports that GP-COMB and a linear-kernel GP reach their performance plateau around 150--200 added points, whereas kriging methods need more points and end slightly less accurate. Kriging still retains value because its variogram-based interpolation yields a more homogeneous spatial distribution of estimated weed cover, whereas kernel methods generalize cover more aggressively across the island.

Load-bearing premise

The evaluation assumes that a random 70/30 split adequately separates training and test sets even though the data are geo-referenced and neighboring points are likely correlated; if that split leaks spatial correlation, the reported interpolation gains may be inflated.

Editorial extensions

If this is right

  • A sugarcane-weed survey of 745 plots can be augmented with about 150--200 GP-interpolated points to reach a lower validation MSE than the original dataset alone.
  • Combined-kernel GP augmentation is not uniformly best for every regressor: quadratic-kernel GP wins for linear and ridge regression, while kriging variants match GP methods for MLP and k-NN, so the augmentation method should be chosen with the downstream model in mind.
  • Because every augmentation method tested improved or matched the baseline, spatial interpolation appears to be a broadly applicable cheap alternative to collecting new field observations for this kind of tabular geo-data.
  • The spatial-coverage result implies a trade-off: if the goal is faithful, smooth coverage maps, kriging may be preferred; if the goal is lowest numeric prediction error, combined-kernel GP is preferred.

Reading between the lines

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

  • The reported gains would be more convincing if evaluated under spatial cross-validation, since the random split may let interpolated smoothness leak between training and test; a buffer-distance split is a direct test.
  • The recipe likely transfers to other costly geo-referenced ecological surveys (other weeds, pests, soil properties), but the relative ranking of kernels and variograms may change with the spatial scale and patchiness of the target.
  • A sharper experiment would compare GP-interpolated points against the same number of randomly drawn synthetic points inside the study region, isolating the contribution of interpolation versus sheer sample size.
  • A multi-species version of this pipeline would be a natural stress test of whether combined-kernel interpolation preserves or erases ecological boundaries between co-occurring weeds.
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 / 6 minor

Summary. The paper studies interpolation-based data augmentation for a geo-referenced weed-cover prediction task on La Réunion, with 745 observations of Commelina benghalensis cover in sugarcane plots. It compares two interpolation families, Gaussian processes with different kernels and kriging with different variograms, as generators of synthetic training points, and evaluates their effect on seven regression algorithms by MSE on a fixed 30% test split. The paper claims that GP-based augmentation, especially with a combined kernel (GP-COMB), significantly improves predictive performance, requiring fewer added points, while kriging variants give slightly lower performance but more homogeneous spatial coverage.

Significance. If the main claim were robust, the paper would offer a simple and potentially useful recipe for augmenting small geo-referenced ecological datasets, and the comparison between kernel-based and variogram-based interpolation is a sensible question. The work has genuine strengths: it uses a real agronomic survey dataset, compares two interpolation families under a common protocol, and includes a spatial-coverage analysis that goes beyond a single performance metric. However, the evaluation protocol has a load-bearing weakness: a single random 70/30 split of spatially autocorrelated geo-referenced data, with no spatial holdout and no repeated-seed variance estimates. The reported gains in Table 3 are therefore not yet credible as evidence of generalization, and the word "significantly" in the abstract is not supported by any significance test. The central claim is defensible in principle, but it needs a substantially strengthened evaluation before publication.

major comments (4)
  1. [Section 3, Protocole; Table 3] The evaluation uses a single random 70/30 split of 745 geo-referenced observations. Because the target variable is spatially continuous and the augmentation methods interpolate using coordinates and covariates, nearby points are likely correlated; under a random split, test points are likely to lie within the correlation range of training points, and augmented points generated in the same region can give regression algorithms near-duplicate information about test locations. The large improvements in Table 3 (e.g., GB from 36.45 to 13.27, RF from 23.83 to 13.35) are consistent with spatial leakage. Please rerun the evaluation with spatial blocking, buffer-distance holdout, or spatial cross-validation, and compare the GP-COMB advantage under that protocol.
  2. [Table 3 and abstract] Only one split is reported for each algorithm, so Table 3 contains no error bars, no repeated-seed variability, and no significance tests. The differences between top methods are often tiny (e.g., KNN 13.17 vs. 13.18 for GP-COMB and CoK-GAU), so the abstract's claim that GP-based methods "significantly improve" performance is not justified by the reported evidence. Report means and standard deviations over multiple random splits and use paired tests or confidence intervals before using the word "significant."
  3. [Section 3, Protocole and Section 4] The paper does not state whether the interpolation models (GP kernels and variograms) used to generate the 200 or 300 augmented points are fit on the training split only or on the full dataset before splitting. If augmented points are produced by models that already saw the test locations, the comparison in Table 3 is circular and the MSE improvements do not measure generalization. Please specify the exact data flow and, if necessary, refit all interpolation models on the training split alone.
  4. [Section 4, 'Nous choisissons pour la suite de ces expérimentations de garder l'algorithme MLP'] The algorithm for the second experiment (performance as a function of number of added points, Figure 1) is selected as the best performer on the same test set used to construct Table 3. This selection on the test set introduces optimism and means the convergence analysis is not an unbiased comparison. Use a separate validation split or nested evaluation to choose the algorithm before studying the augmentation curves.
minor comments (6)
  1. [Table 3] The row label 'BG' should be 'GB' to match the text and the Gradient Boosting abbreviation.
  2. [Section 4, 'Carte de densité'] Figure 3 is referenced before Figure 2 in the text; please renumber the figures and ensure the in-text references match the captions.
  3. [Table 4] The bold entries are described as 'significatives' but no statistical test is reported; either provide significance tests or replace the wording with 'larger differences.'
  4. [Section 3, Protocole] The sentence describing variogram parameter selection as 'en minimisant la MSE' does not say which data are used for that minimization or whether it is cross-validated; specify the fitting procedure.
  5. [Throughout] There are numerous typographical errors and awkward locutions (e.g., 'Proccessus', 'aumgentés', 'algoritme', 'quasiement', 'ormis'); a careful proofreading pass is needed.
  6. [References] Some references are incomplete (e.g., the dataset reference by Fabre-Ferber et al. 2021 has only 'Type: dataset' and no repository or DOI); provide full access information for reproducibility.

Circularity Check

1 steps flagged · score 2.0 of 10

No construction-level circularity; the central augmentation comparison is a genuine empirical benchmark, but the convergence analysis uses an algorithm selected on the same test set, creating a mild self-referential evaluation loop.

  1. other [Section 3 'Protocole' and Section 4 'Performance en fonction du nombre de points rajoutés']
    "L’analyse des performances de l’algorithme le plus performant issu de la première expérimentation, en fonction d’un nombre croissant de points ajoutés {0, 50, . . . ,300} ... Nous choisissons pour la suite de ces expérimentations de garder l’algorithme MLP."

    The convergence experiment, which supports the abstract's claim that GP-COMB and GP-LIN 'require less additional data', is run on MLP, an algorithm chosen because it performed best in the first experiment on the same 30% test set. The resulting MSE-versus-added-points curves are therefore conditional on a model selection made using the test labels; the conclusion that GP-COMB and GP-LIN converge fastest is not an independent prediction but reflects selection on the evaluation set. This is a statistical self-reference rather than an algebraic reduction, so it is mild but real.

full rationale

The paper's central comparison (Table 3) is an empirical benchmark: each interpolation method is fitted to the training data, synthetic points are generated, and the augmented training set is used to train standard regressors evaluated on a held-out test set. This is standard model fitting, not derivation-circularity. The GP and kriging hyperparameters are learned from the data, and the augmented target values are predictions of the interpolation models, but the downstream regressors are separate and the test MSE is a measured outcome, not a quantity forced by construction. The self-citation for the dataset (Fabre-Ferber et al., 2021) is a data provenance reference and is not load-bearing for any methodological claim. The main self-referential element is the choice of MLP for the convergence analysis after observing its performance on the test set; this makes the 'less additional data' conclusion partly selection-dependent. The more serious spatial-leakage risk (random 70/30 split of auto-correlated geo-referenced points) is a correctness/validity threat rather than a circularity, and would require spatial cross-validation to resolve.

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

The paper introduces no new entities. Its central claim relies on standard geostatistical assumptions and on fitted kernel/variogram parameters, whose values are not reported. The most fragile premise is the validity of the random split for spatial data, which is not checked.

free parameters (3)
  • RBF lengthscale sigma = not reported
    Used in the GP-RBF kernel (Eq. 2), optimized by gradient descent on log-likelihood using GPy; affects interpolation and thus all downstream augmentations.
  • Polynomial kernel constant c = not reported
    Used in GP-QUAD kernel (Eq. 3), fit to data; influences synthetic points.
  • Variogram nugget C0, sill C, range a, slope b = not reported
    Parameters of the four variogram models (Eqs. 4-7) are chosen by minimizing MSE via PyKrige; these control kriging weights and synthetic values.
assumptions (3)
  • domain assumption Spatial stationarity: covariance between two points depends only on their distance
    Both kriging and GP interpolation assume spatial correlation decays with distance, as stated in Section 2; if this fails (e.g., due to island terrain), the synthetic points are unreliable.
  • domain assumption The random 70/30 split creates independent training and test sets
    Used implicitly in the protocol (Section 3); this is invalid for spatially autocorrelated data without blocking, threatening the soundness of the evaluation.
  • domain assumption The METEOR service provides accurate auxiliary variables at synthetic locations
    In Section 3, auxiliary variables (temperature, precipitation, luminance) for interpolated points are retrieved from the METEOR service; errors in these retrievals propagate into the augmented dataset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpolation pour l'augmentation de donnees : Application \`a la gestion des adventices de la canne a sucre a la Reunion." pith.science (2026). https://pith.science/paper/M76XZOVO

@misc{pith2026250112400,
  author       = {Pith},
  title        = {Pith review of: Interpolation pour l'augmentation de donnees : Application \`a la gestion des adventices de la canne a sucre a la Reunion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M76XZOVO}},
  note         = {Machine review of arXiv:2501.12400}
}
read the original abstract

Data augmentation is a crucial step in the development of robust supervised learning models, especially when dealing with limited datasets. This study explores interpolation techniques for the augmentation of geo-referenced data, with the aim of predicting the presence of Commelina benghalensis L. in sugarcane plots in La R\'eunion. Given the spatial nature of the data and the high cost of data collection, we evaluated two interpolation approaches: Gaussian processes (GPs) with different kernels and kriging with various variograms. The objectives of this work are threefold: (i) to identify which interpolation methods offer the best predictive performance for various regression algorithms, (ii) to analyze the evolution of performance as a function of the number of observations added, and (iii) to assess the spatial consistency of augmented datasets. The results show that GP-based methods, in particular with combined kernels (GP-COMB), significantly improve the performance of regression algorithms while requiring less additional data. Although kriging shows slightly lower performance, it is distinguished by a more homogeneous spatial coverage, a potential advantage in certain contexts.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 16 canonical work pages

  1. [1]

    Cressie, N. (1990). The origins of kriging. Mathematical geology\/ 22 , 239--252

    work page 1990

  2. [2]

    Cui, L., H. Li, K. Chen, L. Shou, G. Chen (2024). Tabular data augmentation for machine learning: Progress and prospects of embracing generative ai. arXiv preprint arXiv:2407.21523\/

    arXiv 2024

  3. [3]

    Duvenaud, D. (2014). The kernel cookbook: Advice on covariance functions. URL https://www. cs. toronto. edu/duvenaud/cookbook\/

    work page 2014

  4. [4]

    Esnaashari, M. A. H. Damia (2021). Automation of software test data generation using genetic algorithm and reinforcement learning. Expert Systems with Applications\/ 183 , 115446

    work page 2021

  5. [5]

    Schwartz, P

    Fabre-Ferber, F., M. Schwartz, P. Marnotte, T. Le Bourgeois, S. Auzoux (2021). Dataset for the prediction of weediness on sugarcane crop in reunion island. Type: dataset

    work page 2021

  6. [6]

    Baranchuk, I

    Kotelnikov, A., D. Baranchuk, I. Rubachev, A. Babenko (2023). Tabddpm: Modelling tabular data with diffusion models. In International Conference on Machine Learning , pp.\ 17564--17579. PMLR

    work page 2023

  7. [7]

    Ripoche, N

    Laine, E., A. Ripoche, N. Colbach, F. Affholder (2024). Impact of crop-weed competition on yield gap: A field-based approach in sugarcane in r \'e union island. CIRAD

    work page 2024

  8. [8]

    Li, G., Z. Sun, L. Qian, Q. Guo, W. Hu (2021). Rule-based data augmentation for knowledge graph embedding. AI Open\/ 2 , 186--196

    work page 2021

Show all 20 references
  1. [9]

    Naghipourfar, F

    Lotfollahi, M., M. Naghipourfar, F. J. Theis, F. A. Wolf (2020). Conditional out-of-distribution generation for unpaired data using transfer vae. Bioinformatics\/ 36\/ (Supplement\_2), i610--i617

    2020

  2. [10]

    Oliver, M. A., R. Webster, et al. (2015). Basic steps in geostatistics: the variogram and kriging. Technical report, Springer

    2015

  3. [11]

    Ouyang, Y., L. Xie, C. Li, G. Cheng (2023). Missdiff: Training diffusion models on tabular data with missing values. arXiv preprint arXiv:2307.00467\/

    2023 arXiv

  4. [12]

    Varoquaux, A

    Pedregosa, F., G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay (2011). Scikit-learn: Machine learning in P ython. Journal of Machine Lea...

    2011

  5. [13]

    Schwarz, G. (1978). Estimating the dimension of a model. The annals of statistics\/ 6\/ (2), 461--464

    1978

  6. [14]

    Zhang, H

    Wan, Z., Y. Zhang, H. He (2017). Variational autoencoder based synthetic data generation for imbalanced learning. In 2017 IEEE Symposium Series on Computational Intelligence ( SSCI ) , pp.\ 1--7

    2017

  7. [15]

    Williams, C. K. C. E. Rasmussen (2006). Gaussian processes for machine learning , Volume 2. MIT press Cambridge, MA. en

    2006

  8. [16]

    Yu, S., T. Wang, J. Wang (2022). Data Augmentation by Program Transformation . Journal of Systems and Software\/ 190 , 111304

    2022

  9. [17]

    Zha, D., Z. P. Bhat, K. Lai, F. Yang, Z. Jiang, S. Zhong, X. Hu (2023). Data-centric artificial intelligence: A survey. CoRR\/ abs/2303.10158

    2023 arXiv

  10. [18]

    Zheng, S. N. Charoenphakdee (2022). Diffusion models for missing value imputation in tabular data. arXiv preprint arXiv:2210.17128\/

    2022 arXiv

  11. [19]

    Zhu, Y., M. R. Min, A. Kadav, H. P. Graf (2020). S3VAE : Self - Supervised Sequential VAE for Representation Disentanglement and Data Generation . pp.\ 6538--6547

    2020

  12. [20]

    write newline

    " write newline "" before.all 'output.state := FUNCTION article output.bibitem format.authors "author" output.check author format.key output output.year.check new.block format.title "title" output.check new.block crossref missing format.jour.vol output format.article.crossref ...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.