A Sustainable Approach for Assessing Wheat Production in Pakistan Using Machine Learning Algorithms

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study aims to predict wheat production in Pakistan using machine learning algorithms. Given the country's agricultural situation and food security needs, this topic is certainly relevant and important. However, after a thorough review, while the research addresses a crucial issue, it appears to have several areas that could be improved. The main concerns are related to the lack of in-depth analysis, insufficient model justification, and some unclear presentation of results. Here are my detailed comments and suggestions for improvement:

(1) In the introduction, you mention previous studies on crop yield prediction, but the analysis of their limitations is rather brief. I would suggest conducting a more comprehensive review, particularly focusing on studies related to wheat production prediction. It would be helpful to elaborate on what existing research lacks and how your study aims to fill those gaps. For example, when discussing the two main methods for predicting agricultural yield (statistical modeling and process-based growth), a more detailed explanation of why the combination of statistical techniques and machine learning in your study is a better approach would be beneficial.

(2) The identification of the research gap and the highlighting of the novelty of your study could be clearer. Please clearly state what new knowledge or approach your research brings. For instance, you mention using multiple machine learning algorithms, but it would be useful to explain how this combination differs from previous studies and what advantages it offers.

(3) In Section 2.4, you have chosen ANN, MLR, and SVM for the study. However, the justification for this selection seems insufficient. It would be helpful to compare these models with other popular machine learning models, such as random forest or gradient-boosting machines, and explain why these other models were not considered. This will help readers understand the rationality of your model selection and evaluate the comprehensiveness of your research.

(4) Regarding the ANN model in Section 2.4.1, the choice of architecture (5 hidden layers with 10 neurons in each layer) appears somewhat arbitrary. More reasoning behind this choice would be appreciated. Ideally, a hyperparameter tuning process could be conducted to find the optimal architecture for your data and prediction task. Reporting the process and results of this tuning would help justify the selected architecture.

(5) The conclusions in Section 4 are somewhat repetitive and could benefit from greater conciseness. Streamlining the conclusions to clearly present the key findings would be advisable. Avoid repeating information from the results section. Additionally, clearly stating the implications of your research for farmers, policymakers, and investors would be valuable.

(5) The recommendations in Section 4 could be more detailed in terms of implementation. For example, when suggesting that the government launch new projects to develop fertilizers and expand arable land, discussing potential challenges, costs, and timeframes would make your recommendations more practical and useful for relevant stakeholders.

While your research is on an important topic, the manuscript in its current form does not yet meet the publication standards. I would encourage you to carefully consider these comments and conduct a significant revision to address the issues in methodology, data presentation, and result interpretation. A more detailed and rigorous approach will likely be required to make your paper suitable for publication.

Author Response

Comments
This study aims to predict wheat production in Pakistan using machine learning algorithms. Given the country's agricultural situation and food security needs, this topic is certainly relevant and important. However, after a thorough review, while the research addresses a crucial issue, it appears to have several areas that could be improved. The main concerns are related to the lack of in-depth analysis, insufficient model justification, and some unclear presentation of results. Here are my detailed comments and suggestions for improvement:

(1) In the introduction, you mention previous studies on crop yield prediction, but the analysis of their limitations is rather brief. I would suggest conducting a more comprehensive review, particularly focusing on studies related to wheat production prediction. It would be helpful to elaborate on what existing research lacks and how your study aims to fill those gaps. For example, when discussing the two main methods for predicting agricultural yield (statistical modeling and process-based growth), a more detailed explanation of why the combination of statistical techniques and machine learning in your study is a better approach would be beneficial.

(2) The identification of the research gap and the highlighting of the novelty of your study could be clearer. Please clearly state what new knowledge or approach your research brings. For instance, you mention using multiple machine learning algorithms, but it would be useful to explain how this combination differs from previous studies and what advantages it offers.

(3) In Section 2.4, you have chosen ANN, MLR, and SVM for the study. However, the justification for this selection seems insufficient. It would be helpful to compare these models with other popular machine learning models, such as random forest or gradient-boosting machines, and explain why these other models were not considered. This will help readers understand the rationality of your model selection and evaluate the comprehensiveness of your research.

(4) Regarding the ANN model in Section 2.4.1, the choice of architecture (5 hidden layers with 10 neurons in each layer) appears somewhat arbitrary. More reasoning behind this choice would be appreciated. Ideally, a hyperparameter tuning process could be conducted to find the optimal architecture for your data and prediction task. Reporting the process and results of this tuning would help justify the selected architecture.

(5) The conclusions in Section 4 are somewhat repetitive and could benefit from greater conciseness. Streamlining the conclusions to clearly present the key findings would be advisable. Avoid repeating information from the results section. Additionally, clearly stating the implications of your research for farmers, policymakers, and investors would be valuable.

(5) The recommendations in Section 4 could be more detailed in terms of implementation. For example, when suggesting that the government launch new projects to develop fertilizers and expand arable land, discussing potential challenges, costs, and timeframes would make your recommendations more practical and useful for relevant stakeholders.

While your research is on an important topic, the manuscript in its current form does not yet meet the publication standards. I would encourage you to carefully consider these comments and conduct a significant revision to address the issues in methodology, data presentation, and result interpretation. A more detailed and rigorous approach will likely be required to make your paper suitable for publication.

Replies

 Thank you for your valuable suggestions. The issues have been addressed. A brief description is given below:

• The comprehensive review of previous studies regarding wheat production prediction and their limitations has been addressed. Especially, the limitations of the methods for predicting agricultural yields (statistical modeling and process-based growth) have been explained, and we have also explained why the combination of statistical techniques and machine learning in our study would be a better approach for predicting wheat production.
• The potential research gap and clearer novelty have been addressed in the following paragraph, “Comparison of the proposed study with previous studies.”.
• The reason for choosing ANN, MLR, and SVM models for our study is that they are easy, simpler, and widely used algorithms for agriculture yield prediction. The other models, like RF, have some limitations to achieve higher accuracy, which have been explained thoroughly in the introduction section. In addition, the SVM model provided higher accuracy in our study, which is another reason to opt for these techniques.
• We tuned all parameters of the ANN randomly in different ranges, but the model provided the maximum accuracy at the given architectural design (5 hidden layers with 10 neurons in each layer) in our study. Because of using the inbuilt algorithmic functions in MATLAB, the only best-tuned parameters we achieved have been discussed in the results section.

• The repetition of sentences has been removed, and the whole conclusion section has been revised in a streamlined way.

         The all recommendations have been revised in terms of implementation with respect of potential challenges, costs and             timeframes

Reviewer 2 Report

Comments and Suggestions for Authors

In the paper "A sustainable approach for assessing the wheat production in Pakistan using machine learning algorithms", AI-based models were developed using ANN, SVM and MLR to predict Wp. Historical data of the last 60 years in the form of CELF, CA, T, CEGF, CD, AL, FO and RF were used as input parameters, and Wp was used as the output variable. The authors have done an interesting and high-quality work. The regression models they used for forecasting are reliable and simple. It was found that the SVM model works more accurately than the MLR and ANN models.
Comments
However, with a generally positive assessment of the submitted manuscript, we note some problems of this type of research.
• There is no generally accepted set of quality indicators.
• In most cases, there are contradictions between individual quality indicators.
• Different quality indicators correspond to their own best model.
Statistical indicators that assess the quality of a model (significance of coefficients, informativeness, adequacy, stability) are usually used separately, and are not consistent with each other. For example, a model may be inadequate, but informative. Or a model consisting of statistically insignificant coefficients may be informative and adequate. An adequate and informative model may be unstable.
It is also necessary to take into account the problem of multicollinearity, when independent variables are highly correlated with each other, which complicates the interpretation of the influence of individual variables.

Author Response

In the paper "A sustainable approach for assessing the wheat production in Pakistan using machine learning algorithms", AI-based models were developed using ANN, SVM and MLR to predict Wp. Historical data of the last 60 years in the form of CELF, CA, T, CEGF, CD, AL, FO and RF were used as input parameters, and Wp was used as the output variable. The authors have done an interesting and high-quality work. The regression models they used for forecasting are reliable and simple. It was found that the SVM model works more accurately than the MLR and ANN models.

Comments
However, with a generally positive assessment of the submitted manuscript, we note some problems of this type of research.
• There is no generally accepted set of quality indicators.
• In most cases, there are contradictions between individual quality indicators.
• Different quality indicators correspond to their own best model.

Statistical indicators that assess the quality of a model (significance of coefficients, informativeness, adequacy, stability) are usually used separately, and are not consistent with each other. For example, a model may be inadequate, but informative. Or a model consisting of statistically insignificant coefficients may be informative and adequate. An adequate and informative model may beunstable.
It is also necessary to take into account the problem of multicollinearity, when independent variables are highly correlated with each other, which complicates the interpretation of the influence of individual variables.

Replies

Thank you for your valuable feedback. We have highlighted the limitations of model in “limitation and future prospects section”.

All your suggestions are right and useful related to statistical techniques. The contradiction between the statistical analysis with respect of predictors can be resolved in the presence of large and quality dataset.

Reviewer 3 Report

Comments and Suggestions for Authors

I have attached a file with my comments.

Comments for author File: Comments.pdf

Comments on the Quality of English Language

The English language could be refined to more clearly convey the research.

Author Response

Comments

 
This article evaluates three prediction models—two based on artificial intelligence and one

using statistical linear regression—for forecasting wheat production. The study considers key

predictor variables, including wheat area, temperature, rainfall, carbon dioxide emissions

from liquid and gaseous fuel consumption, arable land, credit disbursement, and fertilizer

offtake. Notably, the findings reveal that simpler predictive models outperformed neural

network-based models when data availability was limited.

While the article has strong potential for publication, the following comments should be

addressed.

1. In line 68, what do you mean by "psychological method"? Could you briefly explain it?
2. Please carefully review the various definitions presented in the introduction and ensure

that each definition and fact is properly cited. For example, the statement: “Machine

learning yields crucial results, including predictions that are used to report on future

events (Menchari, 1965)” could not be verified, and its corresponding citation could not

be found. Given the extensive recent literature on machine learning, it would be more

appropriate to cite a recent study rather than a work from 1965.

3. The authors should cite their data sources, whether they are reports, websites, etc. For

example, the article mentions that part of the data was obtained from the Kaggle

platform. Since this platform hosts millions of datasets from various disciplines, the

authors should provide specific links to their data sources.

4. It is recommended to standardize the description of the number of variables. In some

parts of the manuscript, eight predictor variables are mentioned, while in other sections,

such as Section 2.1 and Figure 1, only seven variables are listed.

5. Tables 1 and 2 could be combined into a single table. It is recommended to merge them

for clarity and conciseness.

6. What is the purpose of Table 3? If it is not used later in the analysis, it is recommended to

either remove it or include it as supplementary material.

7. The ANN technique within AI is widely known. A lengthy definition may not be necessary

(Section 2.4.1).

8. The performance evaluation mechanisms for the prediction models, as well as the

evaluation metrics (R², RMSE, MAE) and their equations, should be included in the

methodology section. Move part of section 3.1 to the methodology.

9. Section 3.1 presents and describes the R², RMSE, and MAE values, which are also listed in

Table 1. Since these values are already in the table, refer to the table in the text instead of

repeating them.

10. Please provide the values of the tuned hyperparameters for the ANN and SVM models.

Which hyperparameters were tuned? Additionally, describe the methodology/criteria

used for hyperparameter tuning (e.g., grid search, Monte Carlo, etc.). Also, please

indicate which libraries were used to implement the ML algorithms. Were functions and

libraries used, or was custom code developed? If libraries and public functions were

used, it is recommended to present the code in a repository for future replication.

11. Figure 6 shows that the training was done with data from one range and the testing with

data from another range. This could lead to bias errors. It is suggested that you correct the

training process, or otherwise, discuss this limitation in the study's discussion section

and justify why this approach was used for data training/testing.2

12. Many facts presented at various points in the discussion lack citations. For example: “On

the other hand, ANN usually outperforms the SVM when a large amount of training data is

provided, and ANN processes this information with an adjusted number of hidden layers

until higher accuracy is achieved. The performance and accuracy of these models also

depend on the nature of the dataset, such as whether it is discrete or continuous. In

short, the nature…” does not contain any citation. It is requested that each fact presented

to support your results should be properly cited.

13. It seems that Figure 8 does not provide additional information compared to Figure 7. It is

suggested to remove one of them or include it as supplementary material and consolidate

all the information into a single figure.

14. The criteria for the sensitivity analysis, as well as the equation, should be included in the

methodology section rather than in the results section.

15. The manuscript should clarify that it is not developing or proposing a new machine

learning (ML) methodology but rather applying existing ML models. The development or

proposal of a new ML method typically involves modifications to the algorithm's structure,

the creation of novel optimization strategies, or the development of new techniques,

often involving coding from scratch. In this context, the manuscript does not present such

a contribution. However, the primary contribution of the paper lies in the novel

combination of predictor variables, which is a noteworthy innovation and should be more

thoroughly emphasized. Another important finding is that, using a limited number of

predictors (53 observations), artificial neural networks (ANN) do not outperform simpler

models, such as support vector machines (SVM) and statistical models (MLR), in

predicting wheat yield. This finding is significant, as it challenges the common

assumption that more complex models will automatically yield better results. To

strengthen this aspect, the paper should not only highlight the superior performance of

SVM over the other models but also underscore the underperformance of ANN in such

contexts. Recent studies in other disciplines have reached similar conclusions. For

instance:

• Noa-Yarasca, E.; Osorio Leyton, J.M.; Angerer, J.P. "Deep Learning Model

Effectiveness in Forecasting Limited-Size Aboveground Vegetation Biomass Time

Series: Kenyan Grasslands Case Study." Agronomy 2024, 14, 349.

https://doi.org/10.3390/agronomy14020349

• Jayaprakash, A. "A Comparison of Deep Learning Methods for Time Series Forecasting

with Limited Data." Freie Universität Berlin: Berlin, Germany, 2022.

• Vabalas, A.; Gowen, E.; Poliakoff, E.; Casson, A.J. "Machine Learning Algorithm

Validation with a Limited Sample Size." PLoS ONE 2019, 14, e0224365.

 These studies highlight that, with limited data, there is no guarantee that deep learning

techniques will outperform statistical methods. I would strongly recommend citing these

works to contextualize and support your findings.

16. Align the objectives stated in the last paragraph of the introduction with those mentioned

in the first paragraph of the conclusions. 3

17. It is recommended to improve the conclusions. Conclusions should not repeat all the

values presented in the results but instead highlight the main findings and explain the

reasons behind them.

18. In the conclusions section, please focus on the conclusions drawn from your study and

remove any conclusions that do not stem from your experiments and results. For

example, the statement "These AI-based algorithms are precise, reliable and have higher

dependability rather than the old traditional and econometric models which have lower

accuracy and robustness" should be avoided. Do not generalize about findings that are

not part of your study. Particularly, your study shows that MLR performed better than

ANN.

Replies

Thank you for your valuable feedback. We have addressed the issues you mentioned. Description is given below:

• The term "psychological method," used for process-based models, means we can make a strategy to handle the different agrarian attributes such as soil properties, rainfall, temperature data, solar strength, and management techniques for developing simulated models. In addition, if we know the complete knowledge or accurate guess on behalf of experience about these attributes (how they work properly), the lab experiment can be planned effectively for prediction purposes.
•  The more up-to-date references have been added, and the old reference “(Menchari, 1965)” has been removed.
• The Kaggle link was already provided regarding climate change indicators in Pakistan in the references section.

• The number of variables has been standardized in the whole manuscript description.
• Both tables have been merged into table 1 according to suggested comment.
• Table 3 has been moved to the supplementary file.
• The ANN technique has been summarized.
• The part of section 3.1 has been moved to the methodology.
• The metrics (R², RMSE, MAE) values have been referred as table form in the text.
• We have provided the best-tuned parameters that automatically adjusted during the training process, especially for the SVM technique in MATLAB. Because of using the inbuilt function in MATLAB, some parameters are automatically adjusted, and some parameters, like the number of hidden layers and neurons in the ANN, have been adjusted by ourselves to get higher accuracy.
• The reason for the use of more than one dataset range for training and testing purposes has been discussed clearly in the following section, “Study Limitations and Future Prospects.”.
• All the mentioned facts in the discussion section have been reviewed and properly cited.
• Figure 8 has been removed from the text and included in supplementary material.
• The criteria for sensitivity analysis as well as equations have been transferred to the methodology section.
• The novelty regarding the current study has been explained in the “study limitations and future prospects” section. The limitation of ANN under a lower dataset has also been cited according to the suggested references by the reviewer.
• The paragraph related to study objectives in the introduction has been aligned with those mentioned in the conclusion section.
• The conclusion section has been revised according to the main findings and the reasons behind them.
• The general findings from previous articles has been removed and only the main findings of our study has been explained.

• The queries raised by the reviewer have been addressed. All the changes and amendments suggested by the referee have been incorporated in the manuscript and the manuscript has been revised accordingly and edited linguistically.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I am satisfied with the authors' revisions. They have addressed all the concerns, and the improvements made to the paper are comprehensive and remarkable. The manuscript now shows a substantial enhancement in quality, with the issues related to in - depth analysis, model justification, and result presentation being effectively resolved. There are no remaining major issues or comments from my side.

Reviewer 3 Report

Comments and Suggestions for Authors

The author has addressed all suggested changes and comments. I have no further remarks.

Comments on the Quality of English Language

The clarity and readability of the manuscript could be enhanced with further refinement of the English language. Improving the grammar, phrasing, and overall fluency would help ensure that the research is more effectively communicated.