Pith. sign in

REVIEW 4 major objections 7 minor 17 references

Leveraging Novel Ensemble Learning Techniques and Landsat Multispectral Data for Estimating Olive Yields in Tunisia

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper reports that an automated ensemble on Landsat data estimates olive yields with R² = 0.86, beating earlier models.

desk verdict A clean AutoGluon pipeline whose dominant feature is internally contradictory; the headline R² likely reflects a governorate proxy more than spectral skill. read the letter →

arxiv 2506.06309 v1 pith:TQTUTSG4 submitted 2025-05-26 eess.SP cs.LG

Mohamed Kefi , Tien Dat Pham , Thin Nguyen , Mark G. Tjoelker , Viola Devasirvatham , Kenichi Kashiwagi This is my paper
classification eess.SPcs.LG
keywords oliveyieldestimationLandsat-8OLILandsat-9OLI-2ensemblelearningAutoGluonvegetationindicesremotesensingTunisia
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 olive yields can be estimated accurately and cheaply from free satellite imagery plus a small number of farmer-reported field measurements, using a fully automated ensemble-learning pipeline. It reports that the AutoGluon framework, fed with seven Landsat spectral bands, five vegetation and water indices, and elevation for 192 farms in central Tunisia, predicts 2022/23 yields with $R^2 = 0.8635$ and RMSE $= 1.17$ tons per hectare on Landsat-8, and $R^2 = 0.8378$ and RMSE $= 1.32$ tons per hectare on Landsat-9, under five-fold cross-validation. If the numbers hold, this matters because olive production alternates sharply between years and is economically central to Tunisia; a low-cost, reliable yield estimate would help farmers and policymakers plan in a region where previous machine-learning attempts stayed below $R^2 \approx 0.65$.

What carries the argument

The carrying object is the AutoGluon tabular ensemble: an automated machine-learning framework that trains eight base learners—two neural networks, Random Forest, Extra Trees, CatBoost, XGBoost, LightGBM, and a weighted ensemble—tunes their hyperparameters, and forms a stacked model that learns how to combine their predictions. It matters because the paper's case is precisely that automation, rather than expert feature engineering or a single model, is what pushes olive-yield accuracy past earlier results. The spectral features were fixed before modeling: seven Landsat bands, five indices, and the DEM, all raster-extracted at GPS-surveyed farms and arranged as a 192-row tabular dataset; five-fold cross-validation with automated hyperparameter tuning supplies the reported error metrics.

What would settle it

Re-run the identical AutoGluon pipeline using measured harvest weights from the same 192 farms—weighbridge tickets, cooperative purchase records, or harvester counts—instead of interview responses; if the cross-validated $R^2$ drops materially below 0.86, the paper's accuracy claim depends on the survey ground truth rather than on the satellite features.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that a stacked ensemble, assembled automatically rather than hand-tuned, extracts enough signal from medium-resolution multispectral imagery to predict olive yield at the farm level. The authors combine self-reported yields from 192 olive farms in Kairouan and Sousse with Landsat-8 and Landsat-9 surface reflectance, the vegetation indices NDVI, GNDVI, SAVI, CIgreen, and NDWI, and a digital elevation model, then let the ensemble select and stack base learners. Five-fold cross-validation gives mean $R^2$ values of 0.8635 and 0.8378 for the two sensors, with RMSE values of 1.17 and 1.32 tons per hectare. The paper reports that elevation is the dominant predictor, with higher elevation associated with lower yields, followed by NIR, SWIR, NDVI, and NDWI, and that optical data saturate at yields above roughly 15 tons per hectare.

Load-bearing premise

The load-bearing premise is that the yields reported by farmers during December 2022–January 2023 interviews are accurate enough to serve as ground truth for the satellite model; if those self-reports are biased, inflated, or poorly remembered, the reported $R^2$ and RMSE describe agreement with interview answers, not with true yields.

Editorial extensions

If this is right

  • The same pipeline can be re-applied in other semi-arid olive regions with only a Landsat scene, a DEM, and a small field survey, because all spectral features come from openly available satellite data.
  • Landsat-9 can serve as a drop-in successor to Landsat-8 for olive monitoring, with only a small loss in explained variance, supporting continuity of yield estimates across the two sensors.
  • Elevation-based effects should be included in future yield models for Mediterranean orchards; ignoring them would remove the single most important predictor identified here.
  • Because optical reflectance saturates near 15 tons per hectare, the approach is most trustworthy for typical rainfed and partially irrigated yields, not for extremely high-yield orchards.
  • The reported accuracy comparisons suggest the ensemble approach improves on earlier single-algorithm models, which the paper places below $R^2 \approx 0.65$.

Reading between the lines

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

  • Editorial inference: because the study uses one growing season in two governorates, transfer across years is untested; olive alternation between 'on' and 'off' years means the model should be validated over at least one full cycle before operational use.
  • Editorial inference: the dominance of elevation may partly reflect that Kairouan and Sousse differ systematically in elevation, tree age, and irrigation; with only 192 farms, the model could be learning regional differences rather than a general yield response.
  • Editorial inference: a testable next step the authors do not run is training the same features on a second season or another governorate to check whether the Landsat-8/9 ordering and $R^2$ values are stable, and whether adding radar imagery removes the saturation above 15 tons per hectare.
  • Editorial inference: the ground-truth issue suggests a cheap external check—comparing a subset of farmer-reported yields against olive-oil cooperative intake records for the same farms—to quantify how much of the reported accuracy is survey noise.
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 reports an AutoGluon-based ensemble regression pipeline that estimates olive yield (tons per hectare) in 192 farms in the Tunisian governorates of Kairouan and Sousse using Landsat-8 OLI and Landsat-9 OLI-2 surface reflectance bands, vegetation indices (NDVI, GNDVI, SAVI, CIgreen, NDWI), and elevation (DEM/DSM). With five-fold cross-validation, the authors report mean R² = 0.8635 and RMSE = 1.17 t/ha for Landsat-8, and R² = 0.8378 and RMSE = 1.32 t/ha for Landsat-9. The paper claims a scalable, cost-effective, and accurate method for olive yield estimation.

Significance. If the reported accuracy is substantiated, the paper would provide a useful comparison of Landsat-8 and Landsat-9 for orchard yield estimation, demonstrate the value of automated ensemble stacking (AutoGluon) in a small tabular remote-sensing dataset, and identify NIR, SWIR, NDVI, and NDWI as informative spectral features. The reported R² values exceed those of several prior olive-yield studies cited in the Introduction, so the result is potentially of interest to the agricultural remote-sensing community. However, the support for the central claim is weakened by an internal inconsistency in the dominant elevation feature, a possible governorate-identity confound in the cross-validation design, and the unverified, self-reported nature of the ground-truth yields. These issues must be resolved before the accuracy claim can be accepted.

major comments (4)
  1. [Table 3 vs. Section 3.1] The elevation data are internally contradictory and this is load-bearing. Table 3 lists the average elevation as 46.1 m for Kairouan and 134.7 m for Sousse, while Section 3.1 states that GPS data show Kairouan at approximately 135 m and Sousse at approximately 46 m. Figure 4 identifies the DEM ('Band 13') as the most important feature, with more than 50% importance, and Section 4.2 uses elevation to explain the yield difference between the two governorates. If the DEM values used as model input are the swapped values, the dominant predictor is corrupted and the reported R²/RMSE values cannot be trusted. Please correct the elevation values and the associated text, and rerun the analysis; if the elevation values are correct, clarify the source of the discrepancy.
  2. [Section 2.3.3 and Table 5] The five-fold cross-validation splits farms randomly, but the dataset contains only two governorates with strongly different mean yields (4.6 t/ha in Kairouan vs. 1.1 t/ha in Sousse, Table 3) and the elevation feature differs systematically between them. Because DEM is the dominant feature, the model may achieve high R² by using elevation as a proxy for governorate identity rather than by learning a general spectral–yield relationship. Please report a leave-one-governorate-out (or leave-one-city-out) evaluation, an ablation without the DEM feature, and ideally a prediction check on an independent site or year. Without such analysis, the claim that the model estimates olive yield from spectral data in a transferable way is not established.
  3. [Section 2.2.2] The ground-truth yields are self-reported values obtained through face-to-face interviews with farmers during December 2022 and January 2023. The manuscript does not describe independent yield measurement, harvest records, or administrative verification. If farmers' reports are biased—for example, systematically over- or under-reporting production—the R² and RMSE values in Table 5 do not measure true yield-estimation error. The paper should explicitly state that the target is reported yield, discuss the possible bias and its direction, and temper the abstract's 'accurate' claim accordingly.
  4. [Generalizability] All model development and evaluation use a single season (2022/2023) and a single geographic domain (two governorates in Tunisia). The five-fold CV therefore measures within-sample consistency for that season, not predictive accuracy for new seasons or regions. The Discussion and Conclusion claim potential applicability 'across diverse agricultural regions globally,' but no external validation is provided. Please add a clear limitations paragraph stating that the current evidence is restricted to the surveyed season and sites, and revise the abstract and conclusion to reflect this scope.
minor comments (7)
  1. [Section 2.2.1 and Table 1] Table 1 lists eight bands including 'MIR' and 'SWIR 1' and 'SWIR 2', while Section 2.3.1 says seven bands (B1–B7) were used. Please clarify the exact set of Landsat bands used and ensure the text and table agree.
  2. [Section 2.3.2] The text says 'four vegetation indices including NDVI, GNDVI, SAVI, and Clgreen along with NDWI,' which is five indices. Please correct the count for consistency.
  3. [Section 3.2 and Table 5] There are typos: 'archived' should be 'achieved' and 'Lansat-9' should be 'Landsat-9'. Please proofread these occurrences.
  4. [Equation 7] The definition of ye and ym is ambiguous: please state explicitly that ye is the mean of estimated values and ym is the mean of actual values, or use conventional notation.
  5. [Section 2.3.2] 'Jupiter Notebook' should be 'Jupyter Notebook'.
  6. [Throughout] The abbreviation for the chlorophyll index appears both as 'Clgreen' and 'CIgreen'. Please use one consistent form.
  7. [Figure 4] The label 'Band 13' is used to denote the DEM feature, which is not an actual Landsat band. Please label it directly as 'DEM' to avoid reader confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the reported R² values come from held-out cross-validation and the predictor features are external to the yield measurements.

full rationale

The paper's derivation chain is self-contained and does not reduce to its own inputs. The yield estimates are produced by an AutoGluon ensemble trained on spectral bands, vegetation indices, and DEM values, with performance evaluated by five-fold cross-validation on held-out farm samples; the R² and RMSE values in Table 5 are therefore genuine predictive metrics rather than refitted training values. No fitted parameter is renamed as a prediction, and no target-derived feature is used as a predictor. The DEM elevation is an external geospatial covariate, not a function of yield, so its high importance reflects a statistical association—and possible governorate confounding—rather than definitional circularity. Self-citations (e.g., Kefi et al. 2016, 2022) appear only in background statements about drought and olive farm identification and are not load-bearing for the central yield-estimation claim. The inconsistent elevation values between Table 3 and Section 3.1 and the use of self-reported survey yields are serious data-quality and generalization concerns, but they are not circularity: they do not make the predicted quantity equivalent to the model input by construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on self-reported ground truth, the assumption that the 13 features carry yield information, and the validity of 5-fold CV as a generalization estimate. No invented entities are introduced. The main hidden loads are data accuracy, spatial confounding via elevation, and single-season sampling.

free parameters (1)
  • AutoGluon hyperparameters, model weights, and stacking weights = Not reported
    No hand-set constants are reported; AutoGluon fits model weights, hyperparameters, and stacking weights on the data. The final configuration is not disclosed, and the reported R² depends on it.
assumptions (3)
  • domain assumption Self-reported farmer yields are an accurate ground truth for training and evaluation.
    Section 2.2.2: yields from face-to-face interviews are used as the target variable; no independent measurement or administrative verification is described.
  • domain assumption Spectral bands, vegetation indices, and DEM encode a stable relationship with olive yield across the study area.
    Core model assumption in Section 2.3; the model treats all 13 features as predictors of yield.
  • domain assumption Five-fold cross-validation on 192 plots from two governorates estimates real-world performance.
    Section 2.3.3 and Table 5; there is no independent test year or region, and the sample is small and uneven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Novel Ensemble Learning Techniques and Landsat Multispectral Data for Estimating Olive Yields in Tunisia." pith.science (2026). https://pith.science/paper/TQTUTSG4

@misc{pith2026250606309,
  author       = {Pith},
  title        = {Pith review of: Leveraging Novel Ensemble Learning Techniques and Landsat Multispectral Data for Estimating Olive Yields in Tunisia},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQTUTSG4}},
  note         = {Machine review of arXiv:2506.06309}
}
read the original abstract

Olive production is an important tree crop in Mediterranean climates. However, olive yield varies significantly due to climate change. Accurately estimating yield using remote sensing and machine learning remains a complex challenge. In this study, we developed a streamlined pipeline for olive yield estimation in the Kairouan and Sousse governorates of Tunisia. We extracted features from multispectral reflectance bands, vegetation indices derived from Landsat-8 OLI and Landsat-9 OLI-2 satellite imagery, along with digital elevation model data. These spatial features were combined with ground-based field survey data to form a structured tabular dataset. We then developed an automated ensemble learning framework, implemented using AutoGluon to train and evaluate multiple machine learning models, select optimal combinations through stacking, and generate robust yield predictions using five-fold cross-validation. The results demonstrate strong predictive performance from both sensors, with Landsat-8 OLI achieving R2 = 0.8635 and RMSE = 1.17 tons per ha, and Landsat-9 OLI-2 achieving R2 = 0.8378 and RMSE = 1.32 tons per ha. This study highlights a scalable, cost-effective, and accurate method for olive yield estimation, with potential applicability across diverse agricultural regions globally.

Figures

Figures reproduced from arXiv: 2506.06309 by the authors.

Figure 1
Figure 1. Location map of the study areas [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Flowchart for developing novel ensemble learning model using Landsat data 2.3.1. Satellite image processing Before performing image analysis, preprocessing was applied to Landsat-8 OLI and Landsat￾9 OLI-2 data using ArcGIS Pro V3.2 tools. To accurately identify landscape features, the first step involved creating a single raster from multiple spectral bands. Seven multispectral bands (B1, B2, B3, B4, B5, B6, and B7)… view at source ↗
Figure 3
Figure 3. shows scatterplots of predicted versus actual olive yields across 5-fold CV. The Landsat-8 sensor yielded higher predictive capability and better goodness-of-fit. Both two sensors were able to predict olive yield with precise accuracy, up to 10 tons ha -1 . However, the scatter plots indicate data saturation in the Landsat imagery at yield values exceeding 15 tons ha -1 . (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: presents the variable importance for predicting olive yield in the study areas using the Landsat-8 OLI and Landsat-9 OLI-2 datasets. A similar trend is observed across both datasets. The most important feature across the 5-fold CV is Band 13, representing the DEM, with…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages

  1. [6]

    Accessed date 30 March 2025 Kamiyama, H., Kefi, M., Kashiwagi, K.,

    Productions and exportations database (https://www.internationaloliveoil.org). Accessed date 30 March 2025 Kamiyama, H., Kefi, M., Kashiwagi, K.,

    work page 2025

  2. [10]

    Scalable Computing 23 (2) 59 –79, https://doi.org/10.12694/scpe.v23i2.2025

    Review of Crop Yield Estimation using Machine Learning and Deep Learning Techniques. Scalable Computing 23 (2) 59 –79, https://doi.org/10.12694/scpe.v23i2.2025. Moussaid, A., El Fkihi, S., Zennayi, Y., Lahlou, O., Kassou, I., Bourzeix, F., El Mansouri, L., Imani, Y.,

    work page doi:10.12694/scpe.v23i2.2025 2025

  3. [13]

    Fine -scale ecological and economic assessment of climate change on olive in the Mediterranean Basin reveals winners and losers. Proc. Natl. Acad. Sci. USA 2014, 111, 5598–5603. Stateras, D., Kalivas, D.,

    work page 2014

  4. [14]

    PLoS ONE 20(1): e0311530

    Improving early prediction of crop yield in Spanish olive groves using satellite imagery and machine learning. PLoS ONE 20(1): e0311530. https://doi.org/10.1371/journal. pone.0311530 Rodríguez-Díaz, F., Chacón -Maldonado,A.M., Troncoso -García,A.R., Asencio -Cortés,G.,

    doi:10.1371/journal

  5. [15]

    Rouse,J.W., Haas,R.H., Scheel,J.A., Deering,D.W.,

    Explainable olive grove and grapevine pest forecasting through machine learning-based classification and regression, Results in Engineering, (24), https://doi.org/10.1016/j.rineng.2024.103058. Rouse,J.W., Haas,R.H., Scheel,J.A., Deering,D.W.,

    work page arXiv 2024

  6. [16]

    Crop yield prediction using machine learning: An extensive and systematic literature review, Smart Agricultural Technology, (10),100718, https://doi.org/10.1016/j.atech.2024.100718 Tous, J., Romero, A.,Hermoso, J. F.,

    arXiv 2024

  7. [17]

    Advances in Horticultural Science , 24(1), 43 –52

    New trends in olive orchard design for continuous mechanical harvesting. Advances in Horticultural Science , 24(1), 43 –52. http://www.jstor.org/stable/42882753 Yin, C., Binbin, H., Xingwen, Q., and Liao, Z.,

    work page arXiv

  8. [21]

    H. M. Mamrutha et al. (eds.), Translating Physiological Tools to Augment Crop Breeding. Springer Natur e Singapore Pte Ltd https://doi.org/10.1007/978-981-19-7498-4_21 Marques, P., Pádua, L., Sousa, J.J., Fernandes -Silva, A.,

    work page doi:10.1007/978-981-19-7498-4_21

Show all 17 references
  1. [56]

    https://doi.org/ 10.3390/agronomy11010056 Fraga, H., Guimarães, N., Freitas, T.R., Malheiro, A.C., Santos, J.A

    doi

  2. [350]

    https:// doi.org/10.3390/agronomy12020350 Gao, B.C.,

    doi

  3. [2010]

    Olea europaea L. in the North Mediterranean Basin during the Pleniglacial and the Early–Middle Holocene, Quaternary Science Reviews, Volume 29, Issues 7–8: 952-968, https://doi.org/10.1016/j.quascirev.2009.12.015. Cubillas, J.J., Ramos, M.I., Jurado, J.M., Feito, F.R.,

    2009 doi

  4. [2014]

    Agronomy for Sustainable Development, 34 (3), pp.685 -694

    Better prediction of Mediterranean olive production using pollen-based models. Agronomy for Sustainable Development, 34 (3), pp.685 -694. 10.1007/s13593-013-0198-x. Hal-01234818. 19 Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., P...

    doi

  5. [2020]

    arXiv preprint arXiv:2003.06505

    Autogluon-tabular: Robust and accurate automl for structured data. arXiv preprint arXiv:2003.06505. 17 FAO,

    2003 arXiv

  6. [2021]

    Asian Journal of Agriculture and Rural Development, 11(3), 255 -261

    Irrigation Water Use Efficiency in Olive Trees in Kairouan, Tunisia. Asian Journal of Agriculture and Rural Development, 11(3), 255 -261. 10.18488/journal.ajard.2021.113.255.261 Kamiyama, H., Kashiwagi, K., Kefi, M.,

    2021 doi

  7. [2022]

    Agroecology and Sustainable Food Systems, 46 (7), 1025 –1046

    Adoption factors and structural characteristics of irrigated olive grove agroforestry systems in Central Tunisia. Agroecology and Sustainable Food Systems, 46 (7), 1025 –1046. https://doi.org/10.1080/21683565.2022.2085230 Lobell, D.B., Cassman, K.G., Field, C.B.,

    2022

  8. [2024]

    Addanki, U.K., Maddineni, T., Dhawale, V., Prasad, M.L.M., Reddy, D.N., Jala, J.,

    Crop yield prediction in agriculture: A comprehensive review of machine learning and deep learning approaches, with insights for future research and sustainability, Heliyon, Volume 10, Issue 24, https://doi.org/10.1016/j.heliyon.2024.e40836. Addanki, U.K., Maddineni, T., Dhawa...

    doi

  9. [2025]

    Remote Sensing 17,

    Mapping Vegetation Changes in Mongolian Grasslands (1990–2024) Using Landsat Data and Advanced Machine Learning Algorithm. Remote Sensing 17,

    1990

Pith tools

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