Crop recommendation with machine learning: leveraging environmental and economic factors for optimal crop selection

Silima Marshal DAbreo; Steven Sam

REVIEW 4 major objections 5 minor 5 references

Crop recommendation with machine learning: leveraging environmental and economic factors for optimal crop selection

T0 review · 4 major / 5 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 A Random Forest model with lag variables is the most realistic crop recommendation option tested for India, reaching 83.62% accuracy on a true temporal holdout.

desk verdict A useful but shaky applied-ML paper: the temporal evaluation is a step up from the field's norm, but the top features are unavailable at recommendation time and the headline accuracy is partly an artifact of label leakage. read the letter →

arxiv 2505.21201 v1 pith:DRQNF7GO submitted 2025-05-27 cs.LG

Steven Sam , Silima Marshal DAbreo This is my paper

classification cs.LG

keywords croprecommendationRandomForestSupportVectorMachinelagvariablestemporalvalidationIndianagricultureminimumpriceyieldprediction

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 argues that crop recommendation systems for India should be judged with time-aware evaluation, and that a Random Forest model with lag variables is the best tested option for suggesting one of 19 crops across 15 states. The authors show that standard 10-fold cross-validation gives a misleadingly high RF accuracy of 99.96%, while a chronological split that trains on 2011-2013 and tests on 2014 drops to 78.55%. Adding lag features, meaning past-season weather values and past-year cost and minimum support price values, recovers part of the gap and reaches 83.62% for Random Forest and 74.38% for SVM. The paper concludes that the lag-augmented Random Forest is the most preferred algorithm for the Indian context because it respects temporal order and captures seasonal and inflationary patterns.

What carries the argument

The central mechanism is the lag variable: for temperature, precipitation, humidity and wind speed, the model receives values from up to seven earlier seasons, and for operational cost, fixed cost and minimum support price, values from up to five earlier years. The lags carry the argument by injecting seasonal and inflationary history into the feature set while the chronological 80/20 split keeps the 2014 test year unseen, thereby separating genuine temporal generalisation from the near-perfect but overfit numbers of random cross-validation.

What would settle it

Reserve 2014 strictly for final testing and choose the lag counts on 2013 data alone; if the lag-augmented Random Forest then scores close to the 78.55% no-lag baseline instead of 83.62%, the claimed improvement is a tuning artifact rather than genuine temporal generalisation.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that Random Forest with lag variables is the most preferred algorithm for optimal crop recommendation in the Indian context. Trained on eight environmental factors plus operational cost, fixed cost, total cost and minimum support price, for 19 crops across 15 states, the model is evaluated three ways: random 10-fold cross-validation, a chronological split training on 2011-2013 and testing on 2014, and a lag-variable version of that chronological split. The random split gives RF 99.96% accuracy, which the paper flags as overfitting. The plain chronological split drops RF to 78.55%, and adding lag features recovers some of that loss, raising RF to 83.62% (Kappa 82.01%) and SVM from 71.18% to 74.38%. The paper's conclusion is that the lag-augmented RF is the realistic choice because it maintains temporal order and captures seasonal and inflationary history.

Load-bearing premise

The evaluation assumes the 2014 test year was truly unseen even though the lag counts (7 for seasons, 5 for years) were chosen by watching accuracy improve, and it assumes the farmer already knows the crop's cultivation cost and minimum support price before asking which crop to grow.

Editorial extensions

If this is right

Random 10-fold cross-validation should not be the headline metric for crop recommendation models trained on multi-year data; the paper's own numbers show it can report 99.96% accuracy for a model that drops to 78.55% on a temporal split.
A lag-augmented Random Forest is the model the paper recommends for deployment across the 15 Indian states and 19 crops in the dataset, with 83.62% accuracy and 82.01% Kappa.
Cost and price variables (operational cost, fixed cost and minimum support price) were the strongest predictors in all three RF approaches, so collecting reliable economic data matters at least as much as weather and soil data.
Combining seasonal lags for weather variables with yearly lags for economic variables is a transferable recipe for other time-ordered recommendation problems.
Using Kappa and F1 alongside accuracy gives a more honest comparison across imbalanced crop classes, and the paper shows Kappa consistently sits below accuracy.

Reading between the lines

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

Beyond the paper's own claims, the reported 83.62% is likely optimistic: because Section 3.2.3 says adding lags beyond 7 seasons and 5 years 'failed to improve the accuracies,' the lag counts were selected with knowledge of test-year results, so a second untouched holdout year would give a fairer estimate.
The model cannot be used exactly as described at recommendation time, since a farmer would need to know the season's cultivation cost and minimum support price before deciding which crop to grow; a deployable system would need to forecast those economic variables.
The temporal-split logic generalises beyond India: any crop recommendation study with several years of data should report a chronological split alongside a lag-augmented version, because the random cross-validation number is not a reliable predictor of next-season performance.
A natural next step the paper does not test is ranking crops by predicted profit, such as yield times expected price minus cost, rather than classifying a single crop; the variable-importance results suggest this would shift recommendations further toward economic factors.

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.

Desk Editor's Note

read the letter

The one genuinely useful thing here is the evaluation design. Most crop-recommendation papers report 10-fold CV accuracy in the high 90s and stop. This group actually tests on a later year, watches accuracy drop, and then tries lag features to recover some of the gap. That is the right instinct, and it makes the RF-lag result (83.62%) more credible than the 10-fold number. The dataset itself—19 crops across 15 Indian states with both environmental and economic features—is broader than most district-level studies in this literature, and the data sources are public and named.

The soft spots are real but not all equal. The abstract has a flat error: it says the Lag Variables approach improved performance “compared to the 10-fold cross validation approach,” which is backwards—the lag approach is worse than the 10-fold CV numbers, better than the time-series split. That needs fixing regardless. The SVM numbers also disagree between the abstract (94.71%) and Table 2 (95.73%). Those are fixable.

The more serious issue is the one the stress-test flags. The variable importance plots put Operational Cost, Fixed Cost, and MSP at the top in every approach. Those are crop-specific, year-state-specific values. A farmer choosing a crop doesn't know the cost structure and MSP of a crop they haven't planted yet. So the model is partly recognizing the crop from its own economic fingerprint, not recommending a crop on the basis of suitability. The temporal split does not cure this, because the 2014 test instances still carry their own costs and MSPs. The 83.62% figure is therefore not a clean estimate of recommendation accuracy in the deployment setting. The authors even note in the conclusion that market demand and retail prices should be added, but the deeper problem is that the current top features are not causally available at decision time.

The lag-count selection is a second, smaller issue: Section 3.2.3 says lags beyond 5 years and 7 seasons didn't improve accuracy, which reads like tuning on the test year. One holdout year with no error bars cannot bound generalization. And the paper's own text says Yield was deleted to fix skewness, then the VIF analysis re-uses Yield and reports its VIF—an internal inconsistency the authors noticed but didn't resolve.

Who is this for? Practitioners building crop recommendation systems for India, and researchers in applied ML who want a cautionary example of why random CV overstates performance in temporal settings. The authors deserve a serious referee, but the paper needs revision: correct the abstract and tables, clarify or drop the leaky economic features, specify how lag counts were selected, and add error bars or a second holdout year. I would not cite the headline accuracy as evidence for crop recommendation without those fixes.

Referee Report

4 major / 5 minor

Summary. The paper proposes crop recommendation models for India using Random Forest (RF) and Support Vector Machines (SVM) trained on environmental and economic features for 19 crops across 15 states. The models are evaluated under three protocols: 10-fold cross-validation, time-series split, and a lag-variables approach. The authors report that 10-fold CV yields near-perfect accuracy (RF 99.96%, SVM 94.71%), that temporal split lowers accuracy (RF 78.55%, SVM 71.18%), and that adding lag variables improves accuracy to 83.62% for RF and 74.38% for SVM. The paper concludes that the RF model with lag variables is the most preferred algorithm for optimal crop recommendation in India.

Significance. If the central claim were valid, the paper would make a useful contribution by combining environmental and economic factors in a crop recommendation system and by adopting temporal validation, which is more realistic than the random splits common in this literature. The paper also compares two standard machine learning models and reports a multi-class evaluation with accuracy, Kappa, and F1 scores. However, the evaluation is undermined by a circular feature design: crop-specific cost and minimum support price (MSP) are used as inputs to predict the crop label, and these are unavailable at decision time. The headline numbers are therefore not evidence for crop recommendation as claimed, and the specific 83.62% figure is further compromised by using the test year to select lag counts. As a result, the contribution as stated is not established.

major comments (4)

[§2.3.2, §4.2, Figs. 13a–c] The model uses crop-specific economic features — Operational Cost, Fixed Cost, and MSP — as predictor variables to classify Crop Names. These features are largely determined by the crop itself: each crop has its own cost structure and MSP, and in the deployment setting described in §4.2 the farmer would not know these values before choosing a crop. The variable-importance plots place these features at the top in all approaches, which indicates the classifier can identify the crop from its economic profile rather than recommending a suitable crop. The reported accuracies, including RF-lag 83.62%, are therefore likely inflated by target leakage. The task needs to be redefined so that cost/MSP are either excluded as inputs or supplied for each candidate crop in a scoring/ranking framework; the current classification setup cannot support the paper's recommendation claim.
[§3.2.3] The number of lag variables (7 for seasons, 5 for years) was chosen by monitoring accuracy: the text states that "any further creation of lags beyond these numbers failed to improve the accuracies of the model significantly." This means the 2014 test instances were used during model selection, so the 83.62% accuracy is optimistically biased and is not a clean out-of-sample estimate. The authors should use a nested cross-validation or a separate validation set for lag-count selection before evaluating on a truly held-out test period.
[Abstract and §5] The abstract and conclusion state that the Lag Variables approach improved performance "compared to the 10-fold cross validation approach." This is contradicted by the paper's own numbers: 10-fold CV achieves RF 99.96% and SVM 94.71%, while the lag approach achieves 83.62% and 74.38%. The improvement is actually relative to the Time-series Split approach, not the 10-fold CV approach. This misreporting reverses the direction of the comparison and needs to be corrected.
[§3.2.1, Table 2, Abstract] There are numerical inconsistencies in the reported performance values. Table 2 reports RF accuracy 0.9966 (99.66%) and SVM 0.9573 (95.73%), while the abstract reports RF 99.96% and SVM 94.71%. Section 4.1 also uses 99.96% for RF in Approach 1. The authors should reconcile these numbers and ensure that the abstract, tables, and text report the same values.

minor comments (5)

[§2.1.1–2.1.2] The dataset size is stated as 12,839 instances in §2.1.1 and as 12,749 instances in §2.1.2; the discrepancy should be clarified, since the paper also says 89 missing instances were deleted, which would not account for the difference.
[§2.3.2] The text says the data span "five seasons – kharif, rabi, autumn, summer and winter," but then immediately lists six season categories including "whole year." Please clarify the actual number of distinct season values in the dataset.
[§2.1 and §3.2.1] Table 2 is used twice with different content: once for the economic data dictionary and once for the 10-fold CV performance metrics. The tables should be renumbered to avoid confusion.
[§3.1.3] Figure 8f is referenced for "Operational Cost increase with Years," but the preceding figures are numbered 9a–9g; the figure numbering should be made consistent.
[§2.3.2] The phrase "It aims to stimulate how well the model would perform" should read "simulate." Also, the paper should specify the hyperparameters and random seeds used for RF and SVM to enable reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

Economic inputs are joined by Crop Names, so the RF-lag 'recommendation' partly predicts the label from the label's own cost/MSP profile; lag counts were also tuned on the same test year.

self definitional [Section 2.1.1 (data merging) and Section 4.2 (variable importance)]
"The dataset collected from the three sources were merged together in excel based on three common columns – State Names, Crop Year and Crop Names."

The economic features (Operational Cost, Fixed Cost, Total Cost, MSP) are joined onto each row by Crop Names, the same field used as the prediction target. Section 4.2 reports that these economic factors are consistently placed at the top of variable importance in all RF approaches. Thus the classifier's main signal is a set of values that are properties of the crop label it is asked to output; high accuracy largely means recognizing a crop from its own cost/MSP profile, not identifying an optimal crop from farmer-supplied conditions. The temporal split into 2014 does not repair this, because each 2014 test row still carries that crop's 2014 cost/MSP.
fitted input called prediction [Section 3.2.3 (Lag Variables Approach)]
"We observed that any further creation of lags beyond these numbers failed to improve the accuracies of the model significantly."

The number of lag steps (7 for seasons, 5 for years) was selected by monitoring accuracy while adding lags, and the reported result (RF 83.62%) is evaluated on the same 2014 test partition used for that selection. The improvement over Approach 2 is therefore not a clean out-of-sample prediction; the lag depth is a hyperparameter fitted to the test set, and the reported accuracy already incorporates this selection. The paper gives no separate validation set for choosing the lag counts, so calling the 83.62% figure an independent improved performance overstates what was actually predicted.

full rationale

The paper contains no self-citation chain; its references are external and the modelling pipeline is transparent about the 10-fold cross-validation overfitting concern. However, the central claim that the RF-lag model is the most preferred algorithm for optimal crop recommendation is weakened by two construction-level issues. First, the dataset is merged on Crop Names, and the economic variables attached to each row are crop-specific values; these same features dominate variable importance. That makes the prediction task partly self-definitional: the model is given a crop's own cost and MSP and asked to output that crop's name, so the reported accuracy partly measures identity recognition rather than recommendation quality. The temporal split does not eliminate this leakage because each 2014 test instance still contains its own crop-specific economic values. Second, the lag counts are selected by watching accuracy improve on the same test set used for the reported 83.62% figure, so the improvement over Approach 2 is partially fitted rather than independently predicted. The paper's non-circular contributions include the comparison of temporal splits and the adoption of Kappa and F1 metrics, but the headline accuracy claim does not stand as a fully out-of-sample result.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new entities. Its load-bearing commitments are the lag counts (tuned against test accuracy), the unstated classifier hyperparameters, the join assumption for the three data sources, and the availability of crop-specific economic features before selection. The standard algorithms carry no additional free parameters beyond normal training.

free parameters (3)

Number of season lags for environmental variables = 7
Selected because further lags 'failed to improve the accuracies of the model significantly' (Section 3.2.3), i.e., tuned against the evaluation protocol.
Number of year lags for economic variables = 5
Selected by observing accuracy behaviour in Section 3.2.3; this tuning directly affects the headline 83.62% figure.
SVM and RF hyperparameters = not reported
SVM kernel, cost parameter, and RF tree counts are not stated (Section 2.3), so any tuning that affected results is undisclosed.

assumptions (5)

domain assumption Crop type is the correct classification target for a recommendation system.
The models predict which crop was grown, rather than yield or profit, yet the stated objective is recommending crops with 'maximum yields and profits' (Section 1). The mismatch is not discussed.
domain assumption Merging Kaggle, DES, and Farmer's Portal data on State, Crop Year, and Crop Names preserves correct alignment.
Section 2.1.1 describes a many-to-many merge in Excel; no verification of join cardinality, duplicate matches, or unit consistency is reported.
ad hoc to paper Crop-specific cost and MSP values are known before the crop is selected.
Section 4.2 shows economic variables dominate feature importance, but in the deployment scenario the farmer has not yet chosen the crop, so these inputs would not be available.
domain assumption The 2014 test year is representative of future agricultural conditions.
The central 83.62% accuracy rests on a single holdout year (Section 3.2.2), which cannot capture inter-annual variability in weather, prices, or policy.
standard math Random shuffling in 10-fold cross-validation is a valid evaluation for this temporal dataset.
The authors themselves reject this assumption in Section 3.2.1, noting the near-perfect accuracy 'raises a concern about potential overfitting' due to temporal leakage.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Crop recommendation with machine learning: leveraging environmental and economic factors for optimal crop selection." pith.science (2026). https://pith.science/paper/DRQNF7GO

@misc{pith2026250521201,
  author       = {Pith},
  title        = {Pith review of: Crop recommendation with machine learning: leveraging environmental and economic factors for optimal crop selection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DRQNF7GO}},
  note         = {Machine review of arXiv:2505.21201}
}

read the original abstract

Agriculture constitutes a primary source of food production, economic growth and employment in India, but the sector is confronted with low farm productivity and yields aggravated by increased pressure on natural resources and adverse climate change variability. Efforts involving green revolution, land irrigations, improved seeds and organic farming have yielded suboptimal outcomes. The adoption of computational tools like crop recommendation systems offers a new way to provide insights and help farmers tackle low productivity. However, most agricultural recommendation systems in India focus narrowly on environmental factors and regions, limiting accurate predictions of high-yield, profitable crops. This study uses environmental and economic factors with 19 crops across 15 states to develop and evaluate Random Forest and SVM models using 10-fold Cross Validation, Time-series Split, and Lag Variables. The 10-fold cross validation showed high accuracy (RF: 99.96%, SVM: 94.71%) but raised overfitting concerns. Introducing temporal order, better reflecting real-world conditions, reduced performance (RF: 78.55%, SVM: 71.18%) in the Time-series Split.To further increase the model accuracy while maintaining the temporal order, the Lag Variables approach was employed, which resulted in improved performance (RF: 83.62%, SVM: 74.38%) compared to the 10-fold cross validation approach. Overall, the models in the Time-series Split and Lag Variable Approaches offer practical insights by handling temporal dependencies and enhancing its adaptability to changing agricultural conditions over time. Consequently, the study shows the Random Forest model developed based on the Lag Variables as the most preferred algorithm for optimal crop recommendation in the Indian context.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

[1]

Introduction Agriculture is not only fundamental for food production but also constitutes a primary source for economic growth, employment and improvement of the wellbeing of many people globally. For example, the World Bank reports that agriculture constitutes about 4 % of the world's total gross domestic product (GDP), and in certain least developed nat...

work page 2020
[2]

Method This section describes the approaches for: (i) data collection and pre-processing, (ii) exploratory data analysis (iii) modelling, and (iv) performance evaluation. 2.1 Data Collection and pre-processing We combined environmental and economic input parameters to develop and evaluate the accuracy of two machine -learning models (Random Forest and Sup...

work page 2023
[4]

It exhibits a remarkable accuracy of 99.96% in Approach 1 followed by 78.55 % in Approach 2 and 83.62% in Approach 3

Discussion 4.1 Model performance In all the three approaches, RF outperforms the SVM algorithm confirming its suitability in building crop recommendation models. It exhibits a remarkable accuracy of 99.96% in Approach 1 followed by 78.55 % in Approach 2 and 83.62% in Approach 3. This outcome aligns with the findings reported by Geetha et al (2020) and Lat...

work page 2020
[5]

Conclusion This work proposed comprehensive crop recommendati on models to help farmers in India optimise their crop selection decisions and reduce risk of crop failure while maximising profitability. Currently, most existing crop recommendation systems primarily rely on just environmental factors, cover limited geographical areas and recommend only a han...

work page 2016
[2014]

The split is a standard approach in time series analysis. It aims to stimulate how well the model would perform in predicting future values based on patterns observed in historical data while avoiding the pitfalls of data leakage and ensuring that the temporal order is preserved (Talagala et al., 2018). In the Lag variables Approach 3, along with the trai...

work page 2023

Pith tools

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

[1] [1]

Introduction Agriculture is not only fundamental for food production but also constitutes a primary source for economic growth, employment and improvement of the wellbeing of many people globally. For example, the World Bank reports that agriculture constitutes about 4 % of the world's total gross domestic product (GDP), and in certain least developed nat...

work page 2020

[2] [2]

Method This section describes the approaches for: (i) data collection and pre-processing, (ii) exploratory data analysis (iii) modelling, and (iv) performance evaluation. 2.1 Data Collection and pre-processing We combined environmental and economic input parameters to develop and evaluate the accuracy of two machine -learning models (Random Forest and Sup...

work page 2023

[3] [4]

It exhibits a remarkable accuracy of 99.96% in Approach 1 followed by 78.55 % in Approach 2 and 83.62% in Approach 3

Discussion 4.1 Model performance In all the three approaches, RF outperforms the SVM algorithm confirming its suitability in building crop recommendation models. It exhibits a remarkable accuracy of 99.96% in Approach 1 followed by 78.55 % in Approach 2 and 83.62% in Approach 3. This outcome aligns with the findings reported by Geetha et al (2020) and Lat...

work page 2020

[4] [5]

Conclusion This work proposed comprehensive crop recommendati on models to help farmers in India optimise their crop selection decisions and reduce risk of crop failure while maximising profitability. Currently, most existing crop recommendation systems primarily rely on just environmental factors, cover limited geographical areas and recommend only a han...

work page 2016

[5] [2014]

The split is a standard approach in time series analysis. It aims to stimulate how well the model would perform in predicting future values based on patterns observed in historical data while avoiding the pitfalls of data leakage and ensuring that the temporal order is preserved (Talagala et al., 2018). In the Lag variables Approach 3, along with the trai...

work page 2023