A Machine Learning-Based Assessment of Maize Silage Dry Matter Losses by Net-Bags Buried in Farm Bunker Silos

Round 1

Reviewer 1 Report

The manuscript presents the results on the evaluation of maize silage dry matter losses (DML) using the tree algorithm. Generally, the manuscript is presented in a well-structured manner. Most of the methods and results are clearly described.

Information on machine learning and its applications should be supplemented in the Introduction.

lines 130-136: Please add how many cases the dataset included.

Would it be possible to use other machine learning techniques or more cases to increase the correctness of the models?

Suggestions for future studies should be extended.

Author Response

Reviewer 1 (Modifications on the text: green highlighted)       
General comments - The manuscript presents the results on the evaluation of maize silage dry matter losses (DML) using the tree algorithm. Generally, the manuscript is presented in a well-structured manner. Most of the methods and results are clearly described. Information on machine learning and its applications should be supplemented in the Introduction.

Author’s Reply – Following the Reviewer’s remarks we have added information on machine learning and its application in the introduction (see L. 54-59) and discussion (see L. 284-288). A new reference has been added in the manuscript (n. 14)

English language and style are fine/minor spell check required.

Author’s Reply – The authors have revised the grammatical inconsistencies and improved the style. See the tracked and/or blue highlighted changes in the manuscript.

 

Specific comments:

1. Lines 130-136: Please add how many cases the dataset included.

Author’s Reply – The authors have added the sampling size of the dataset (see L. 137).

 

2. Would it be possible to use other machine learning techniques or more cases to increase the correctness of the models?

Author’s Reply – The authors carried out also other machine learning techniques such as a random forest feature selection followed by a factorial discriminant analysis but the most predictive chemical variables (NDF, lignin, water-soluble carbohydrates) and their discriminant capacity were the same (or even lower than) of the adopted recursive partitioning decision tree. Indeed, the use of a larger dataset might improve the accuracy of the proposed predictive models to assess the dry matter losses of ensiled whole-plant maize. However, the choice of sampling size was a trade-off between the experimental conditions and the robustness of the algorithms. It was quite difficult to find a pool of (dairy) farms having a high agronomic potential for maize cultivation and adopting a set of good practices of ensiling whole-plant maize in bunker silos, so that the variability in terms of silage DML could be attributed utmost to the investigated variables (i.e., proximate composition of fresh forage, silage density and porosity).

 

3. Suggestions for future studies should be extended.

Author’s Reply – The authors have extended the suggestions (see L. 294-297 and 315-317).

 

Reviewer 2 Report

 

This study focused on the estimation of dry matter losses of maize silage. The outcome of this research work is of great practical value and it is very important for maize silage production. But there are still some questions to answer.

 

Line 46: Write more about other practice methods for silages losses detection.

 

Line 63: Complete the FAO range for tested maize hybrids.

 

Line 127 to 129: Complete here why you used these ranges of DML. Is it based on any reference?

 

It is unusual that higher dry matter of fresh maize leads to higher DM losses. It means that farmers were not able to stomp silage enough. But when it is done for enough long time so it should not be problem when maize DM is between 35 and 40 % of DM. Normal are DM losses bigger for wet silages. Please discuss it more in manuscript.

 

References 24 and 29 should be completed.

Author Response

Reviewer 2 (Modifications on the text: yellow highlighted)      
General comments - This study focused on the estimation of dry matter losses of maize silage. The outcome of this research work is of great practical value and it is very important for maize silage production.

Author’s Reply – The authors thank for the positive comment.

 

But there are still some questions to answer.

1. Line 46: Write more about other practice methods for silages losses detection.

 Author’s Reply – The authors have improved this paragraph (see L. 46-50).

 

2. Line 63: Complete the FAO range for tested maize hybrids.

  Author’s Reply – The authors add the information (see L. 71-72).

 

3. Line 127 to 129: Complete here why you used these ranges of DML. Is it based on any reference?

Author’s Reply – As stated, the authors chose these range of DML according to average and s.d. (L. 132). However, a reference (Köhler et al., 2019) has been added to support this DML classification criterium (see L. 130-131 and reference n. 1).

 

4. It is unusual that higher dry matter of fresh maize leads to higher DM losses. It means that farmers were not able to stomp silage enough. But when it is done for enough long time so it should not be problem when maize DM is between 35 and 40 % of DM. Normal are DM losses bigger for wet silages. Please discuss it more in manuscript.

Author’s Reply – The authors agree with the reviewer that several factors affect the DML, including DM at harvest and porosity; above all, the last two depend on the intensity of the packing. However, as suggested by the reviewer, an adequate intensity is not always warranted. This fact could be especially actual in the Italian context, where the harvesting and the silo filling are committed to external operators, whose convenience is to complete the filling of the bunker silo as quickly as possible. Our findings confirm a low DM density if compared to the literature, and under this perspective, the results are in line with the reported effects for such conditions. We added a comment in the manuscript supported by further literature (see L. 243-261; references n. 10, 30, 31).

 

5. References 24 and 29 should be completed.

Author’s Reply – The Authors completed the references (see references n. 25 and 33 in the revised version)

 

Reviewer 3 Report

Review comments on “A Machine Learning-based Assessment of Maize Silage Dry Matter Losses by Net-bags Buried in Farm Bunker Silos” by Severino Segato et al.

 

This paper presents suggests that the DM content of fresh WPM strongly affects the 268

DML across the ensiling process due to the different silage densities achieved in the 269

bunker silos.

 

My main general comments are as below:

- The authors didn’t provide a comparison of the performances on training and testing sets. The authors should investigate experimentally the overfitting of the proposed models.

It is not clear in the formula (1) which variables are the regression coefficients, why notation b is regression coefficients?
The authors did not investigate different regression models for proximate Composition, Fermentative Profile, Density, Porosity, and Dry Matter Losses?
Why were the DML values split into three quantitative classes?
The authors did not apply other methods of machine learning for the prediction of a low level of DML (L-class) along the ensiling process of WPM?
The authors didn’t provide a comparison of the performances on training and testing sets. The authors should investigate experimentally the overfitting of the proposed model.
Is there enough accuracy for the proposed model? Very small Original farm-derived dataset (n = 36) was used in this work.

 

 

Author Response

Reviewer 3 (Modifications on the text: grey highlighted)         
This paper presents suggests that the DM content of fresh WPM strongly affects the DML across the ensiling process due to the different silage densities achieved in the bunker silos.

 Author’s Reply – The Authors confirmed that this was the main goal.

My main general comments are as below:

1. The authors didn’t provide a comparison of the performances on training and testing sets. The authors should investigate experimentally the overfitting of the proposed models.

Author’s Reply – The authors confirmed that the experimental study was planned as a trade-off between the experimental conditions and the robustness of the proposed algorithms. As reported to the reviewer 1, it was quite difficult to find a pool of (dairy) farms having an high agronomic potential for maize cultivation and adopting a set of good practices of ensiling whole-plant maize in bunker silos, so that the variability in terms of silage DML could be attributed utmost to the investigated variables (i.e., chemical composition of fresh forage, silage density and porosity). Thus, the sampling size of the trial did not allow to split the experimental dataset in training and testing set. The machine learning part in this work consists of applying 10 folds cross-validation to optimise the final decision tree. Based on the performance of this approaches we are quite sure to avoid most part of overfitting of the model. Given the limited number of samples, the split of the dataset into train and validation is quite questionable. As stated in the conclusion, the results of this study have to be integrated with further similar trials in the future to verify the accuracy and repeatability of the tested models (see L. 315-317).

2. It is not clear in the formula (1) which variables are the regression coefficients, why notation b is regression coefficients?

Author’s Reply – The authors clarified the variables of formula 1 changing bulk density in silage density and using capital letters for the acronyms (see L. 108-110). Formula 1 is the equation used to calculate porosity as suggested by Richard et al. 2004 (reference 18), isn’t a regression model.

3. The authors did not investigate different regression models for proximate Composition, Fermentative Profile, Density, Porosity, and Dry Matter Losses?

Author’s Reply – The authors confirm that the main objective of the study was to define a predictive model able to estimate the effect of the chemical composition of fresh forage on the DML occurring during its ensiling process. Because of dry matter and chemical composition of fresh forage affect its bunker silo density and porosity, these physical traits of silage were also investigated to assess their impact on DML. The effect of the fermentative profile of silage on its DML was not considered as do not fall within the predictive challenging of the study. 

4. Why were the DML values split into three quantitative classes?

Author’s Reply – The Authors split the DML dataset in three quantitative classes to carry out the recursive partitioning&classification tree. Since the variable under investigation is normally distributed, the split into 3 classes (or 5 classes) makes it possible to better approximate the profile of the distribution itself (about 50% of the data in the middle class - mean +/- 1/2 ds, and 25% in the extreme tails).

5. The authors did not apply other methods of machine learning for the prediction of a low level of DML (L-class) along the ensiling process of WPM?

Author’s Reply – Tree-based methods are simple and useful for interpretation among the supervised machine learning approaches. If the relationship between the features and the response is well approximated by a linear model then a traditional approach such as linear regression will likely work well. If instead there is a complex relationship between the features and the response then decision trees may outperform classical approaches (James et al., 2017. An introduction to statistical learning with application in R, ©Springer). However, this approach allowed us to efficiently separate all 3 classes (including class L) according to the main predictors

6. The authors didn’t provide a comparison of the performances on training and testing sets. The authors should investigate experimentally the overfitting of the proposed model.

Author’s Reply – As already replied, the authors did not provide a validation of the model in a testing set. Based on the performance of 10 folds cross-validation, it was avoided most part of overfitting of the model. According to the sampling size of the investigated dataset, the split into training and validation sets is quite questionable. As reported in the discussion and conclusion, we will recommend further trial to confirm the outcomes of this study.

7. Is there enough accuracy for the proposed model? Very small Original farm-derived dataset (n = 36) was used in this work.

Author’s Reply – As already replied, the investigated dataset does not allow an application of machine learning methods specific for large dataset. Since the relationship between the features (chemical variables) and the response (DML) is approximated by a linear model both the linear regression and the decision tree are enough accurate despite a relative low sampling size. Again, we recommend further trial to confirm the outcomes of this study; a sentence was included in the discussion section (see L. 315-317).