PaperlayerView exact revision on arXiv

Source-linked AI summary

Crop Yield Prediction Using Regression Models: A Comparative Evaluation of Model Performance

Saptak Das, Arijit Sarkar, Soumik Neogi, Suborno Naskar

arXiv:2610.37022v1cs.LGcs.AI

TL;DR

Crop-yield prediction can support agricultural planning and resource allocation, but model performance depends on data and evaluation strategy. This study compares five regression models across two train–test splits, finding similar test performance among linear, ridge, and lasso regression, with lasso marginally lowest on test errors in the 70:30 experiment.

  • Problem

    Crop-yield prediction supports agricultural planning and resource allocation, while model performance is highly data-dependent and influenced by training and evaluation strategies.

    Source: §1 · p. 1

  • Method

    The study evaluates five regression models using 80:20 and 70:30 train–test splits and R2, MSE, and RMSE.

    Source: §1 · p. 2

  • Results

    LR, ridge, and lasso had similar test performance, with R2 values close to 0.913 and RMSE values of approximately 0.501; lasso had the lowest reported test MSE and RMSE in the 70:30 experiment.

    Source: §6 · p. 9, §5.4 · p. 9

  • Takeaways & Limitations

    The evaluated linear model provides a strong baseline for this dataset.

    Source: §6 · p. 9

    Because models used different configurations, results compare evaluated configurations rather than establishing overall algorithm or train–test-ratio superiority.

    Source: §6 · p. 9

Abstract

from arXiv · show

Reliable crop-yield prediction can support agricultural planning and resource allocation. This study evaluates five regression models for predicting crop yield from environmental, crop-related, and agricultural management variables. The dataset initially contained 1,000,000 observations and ten variables, with 231 observations having negative target values removed during preprocessing. Categorical variables were represented using one-hot encoding, and Boolean variables were converted into binary values. The models considered were multiple linear regression (LR), random-forest regression (RF), decision-tree regression (DT), ridge regression, and lasso regression. Their performance was evaluated using 80:20 and 70:30 training-testing splits based on the coefficient of determination (R-squared), mean squared error (MSE), and root mean squared error (RMSE). Applying an 80:20 split, LR achieved a test R-squared of approximately 0.9130 and a test RMSE of 0.5008. Applying a 70:30 split, lasso regression yielded the highest test R-squared (approximately 0.9131) and the lowest test RMSE (approximately 0.5008), with performance almost identical to LR and ridge regression. DT also showed improved test performance under the 70:30 split, whereas RF exhibited a substantial gap between training and testing performance. These results indicate that the evaluated linear regression model provides competitive baselines for this dataset. Further experimentation, including validation testing, will help determine the models' performance and predictability.

1 Introduction

The study compares five regression approaches for crop-yield prediction and examines how data representation and train–test strategy relate to model performance. Its results report similar test R2 values for linear, ridge, and lasso regression, with configuration-specific differences across models.

1 Introduction

  • The dataset combines regional, soil, crop, weather, and management predictors to estimate yield in tonnes per hectare.Observations with negative target values were removed as physically meaningless for crop yield in this study.Source: §1 · p. 2
  • One-hot encoding yields 24 features, while dropping one indicator for each categorical variable reduces the representation to 20 and affects coefficient specification and interpretation.The representation may also affect predictive power.Source: §1 · p. 2
  • The study evaluates five regression models under 80:20 and 70:30 train–test splits, using R2, MSE, and RMSE.The models are multiple linear regression, random-forest regression, decision-tree regression, ridge regression, and lasso regression.Source: §1 · p. 2
  • LR, ridge, and lasso regression achieved test R2 values of approximately 0.9130 for both split ratios; lasso had the lowest test MSE and RMSE under 70:30.The reported conclusions apply to evaluated configurations, not definitive algorithm or split superiority.Source: §1 · p. 2
  • The study’s contributions include crop-yield data preprocessing, comparison of model performance across splits, and analysis of linear-model similarity and the random-forest generalization gap.The listed comparison uses training and testing performance.Source: §1 · p. 2

2 Dataset Description

The study uses a supplied crop-yield dataset with one million rows and ten variables. It removes observations with negative yield targets before analysis.

2 Dataset Description

  • The raw dataset contains 1,000,000 rows and 10 variables.The supplied dataset is identified as Agriculture Crop Yield by Samuel Oti Attakorah [7].Source: §2 · p. 2
  • Removing 231 observations with negative Yield_tons_per_hectare values leaves 999,769 observations for analysis.The negative values were excluded because they are not physically meaningful as crop yields in this study.Source: §2 · p. 3
  • Across the remaining 999,769 observations, Yield_tons_per_hectare ranges from approximately 0.00041 to 9.963 tonnes per hectare.The dataset contains no missing values in its ten columns.Source: §2 · p. 3

3 Related Work

Related work situates crop-yield estimation within broader agricultural machine-learning applications and motivates attention to reported performance factors. The paper compares established linear and tree-based regression approaches.

3 Related Work

  • Reviews cover machine-learning applications in crop and livestock management, disease detection, and yield estimation, including crop-yield methods and factors affecting reported performance.These reviews motivate consideration of data quality, predictor construction, validation strategy, and generalisability.Source: §3 · pp. 3–4
  • Decision trees partition feature space, while random forests aggregate tree predictions to reduce variance and improve robustness.Source: §3 · p. 4
  • Ridge penalizes squared coefficient magnitude, whereas lasso uses an absolute-value penalty that can shrink coefficients to zero.Linear regression provides an interpretable linear baseline.Source: §3 · p. 4

4 Methodology

The methodology preprocesses crop-yield data, compares five regression methods under two train–test ratios, and evaluates predictive performance with three metrics. Hyperparameter selection uses training data while reserving test data for final evaluation.

4.1 Data preprocessing

  • Categorical variables are one-hot encoded and Boolean variables converted to binary values; the full design has 24 predictors, while the reference-dropped design has 20.The target is separated from predictors, and 231 negative-target observations are removed from the one-million-row dataset.Source: §4.1 · p. 4

4.2 Machine-learning models

  • The five methods are multiple linear regression, random-forest regression, decision-tree regression, ridge regression, and lasso regression.Source: §4.2 · p. 5

4.3 Training and hyperparameter-search protocol

  • Experiments use 80:20 and 70:30 splits, compare tuned and untuned configurations, and retain the better-performing configuration for reporting.Search uses training data only, with the test set held out for final evaluation; reported results are selected configurations rather than a uniform tuned-versus-untuned comparison.Source: §4.3 · p. 5

4.4 Evaluation metrics

  • R2, MSE, and RMSE are calculated for each algorithm; higher test R2 and lower test MSE/RMSE indicate better performance on the same test observations.The metric definitions use observed and predicted yields and the observed mean.Source: §4.4 · p. 5

5 Experimental Study

Across the two train–test splits, linear regression, ridge, and lasso had similar test performance; RF showed a substantial training–testing gap, while DT improved under the 70:30 split. The experiments compared five regression models across split and configuration choices.

5.1 Experimental Configurations

  • The study compared five regression models using 80:20 and 70:30 splits, evaluating training and test R2, MSE, and RMSE across configurations that differed in predictors and tuning.Source: §5.1 · pp. 5–6, Table 2 · p. 6

5.2 Results for the 80:20 Train–Test Split

  • Under the 80:20 split, LR, ridge, and lasso had similar test R2 of approximately 0.9130, while RF’s test R2 fell to approximately 0.9077 despite training R2 of approximately 0.9871.Source: §5.2 · p. 6, Table 3 · p. 7

5.3 Results for the 70:30 Train–Test Split

  • Under the 70:30 split, LR, ridge, and lasso test R2 values were close to 0.9131, DT reached approximately 0.9120, and RF retained a training–testing gap.Source: §5.3 · pp. 7–8, Table 5 · p. 8

5.4 Comparative Discussion

  • Test R2 values for LR, ridge, and lasso stayed near 0.913 across splits; lasso had the lowest displayed test MSE and RMSE in the 70:30 experiment, with only a marginal advantage.Source: §5.4 · pp. 8–9, Table 5 · p. 8
  • RF’s training scores substantially exceeded its testing scores, while DT had a smaller training–testing gap and higher test R2 under the 70:30 split.Source: §5.4 · p. 9

6 Conclusion and Future Scope

LR, ridge, and lasso showed similar test performance, while RF had a substantial training–testing gap. The authors describe the results as a dataset-specific baseline and call for further validation and generalization testing.

6 Conclusion and Future Scope

  • R2 values close to 0.913 and RMSE values of approximately 0.501 were reported for LR, ridge, and lasso; lasso had the lowest test MSE and RMSE at 70:30, DT was competitive, and RF overfit.The reported differences among the three linear models were small.Source: §6 · p. 9
  • The findings support a strong linear-model baseline for this dataset, but varied model configurations prevent claims of overall algorithm or train–test-ratio superiority.The authors characterize the comparison as one among configurations rather than algorithms or split ratios in general.Source: §6 · p. 9
  • Future work proposes controlled partitions, systematic tuning, repeated splits or cross-validation, and evaluation on independently collected data across growing seasons and locales.The authors also propose testing boosting methods and assessing feature importance.Source: §6 · p. 9

7 Additional Experimental Results (Appendix)

The appendix presents regression-performance tables for alternative train–test splits, feature counts, and hyperparameter-tuning configurations; the supplied captions identify their settings, not their comparative outcomes.

7 Additional Experimental Results (Appendix)

Loading 2610.37022v1…