Improvement of Wheat Grain Yield Prediction Model Performance Based on Stacking Technique

Round 1

Reviewer 1 Report

The submitted manuscript reports a new stacking regression method that improves the performance of the winter wheat grain yield (GY) prediction model. The authors reviewed in the state-of-the-art the different approaches previously reported in the available literature to evaluate GY. The acquisition of data with near-ground remote sensing technology improves the conventional GY measurement method and overcomes the limitations of the application of satellite-based platforms. Hyperspectral reflectance data has proven to be effective as an input variable for regression models to determine GY. Nevertheless, saturation affects GY accuracy in prediction models constructed with spectral data used as a single value. Leaf area index (LAI) can be an indicator to improve both the accuracy and robustness of prediction models when they use ground spectral data. However, it is necessary to acquire LAI, not of a single-stage but also multiple growth stages, to make effective and accurate predictions. Both spectral and LAI data acquired simultaneously can be adjusted using an elastic network (ENET) algorithm to build individual-growth-stage regression models. Finally, the stacking method allows for combining the results of the different growth stages, obtaining a robust prediction model with improved accuracy on the estimation of the GY prediction.

 

The topic of this manuscript is of interest, outdoor crops represent an essential part of the human diet, and both naturally occurring uncontrollable factors and the complex interactions between them affect their growth. Therefore, the availability of statistical methods that can accurately predict production is essential to agricultural science and even more in the current context of climate change. Besides reporting valuable data for winter wheat, the study is valid as a reference model to other crops. The experimental data were obtained from sixteen different treatments done in triplicate to allow for statistical variability, using different levels for independent variables as fertilizer, irrigation, and crop varieties. The acquired hyperspectral data was simplified to fifteen vegetation indices used with LAI in a well-designed statistical treatment of results that support the conclusions drawn from the study.

 

Therefore, after reviewing the manuscript, I recommend the Editor´s acceptance for publication, having only minor appointment and questions for the authors.

 

• The authors report 16 treatments with four levels of nitrogen fertilizer, three levels of water irrigation, and two crop varieties. Please include a table that summarizes the treatment combinations tested.
• Each treatment was done in triplicate to allow for statistical variability. Information about the mean and the covariance of each treatment is of interest.
• The accuracy obtained in GY prediction at the jointing stage is low. The performance of the stacking method would be improved by avoiding including data from this early stage of the wheat growth?

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

The work concerns the :„Improvement of Wheat Grain Yield Prediction Model Performance Based on Stacking Technique”.
 
The methodological description and the results interpretation  is detailed and comprehensive.  The valuable side of the work is development of methodology of stacking regression model for predicting grain, which combines the usage of spectral information derived from every 4 stages of wheat growth. Such an approach to spectral data analysis may be the future in   application the remote sensing data  for yield forecasting practice.
 
The use of a large number (15 spectral) of indicators to describe 48 objects (yields) employed in model development is its  drawback. Even though the fact that the authors had used analyzes to optimize the selection of data for analysis, the lack of model verification on independent data (another years, others objects) is its another weakness.  
 In the present state, the model does describe the analyzed objects, but it does not seem to be reliable for general estimating yields in practice.
I believe, that verifying the model on independent data would significantly increase the value of the work
 
I would suggest authors to consider developing a new model that will take into account Vegetation Indices available from satellite data,  at  the studied phases of the wheat development.
Then, using Vegetation Indices from satellite measurements, validate the developed model on independent data (yields).
 
 
Requires clarification and correction:
1.     
There are LAI data in table 3, but are given in wrong units 
(t/ha) 
Since LAI is defined as the ratio of the leaf area of the plant  (m2) to the area of the ground the plant occupies (m2) as a factor is  expressed in  m2/ m2 . 
 
In general, LAI is strongly correlated with the yield, which was also emphasized by the authors in the introduction (verses 52-71). 
The data in table 2 "Main statistical parameters of grain yield from different growing seasons" shows different, even strange observations in two following seasons.  Therefore, the authors owe us an explanation or verification of the data (grain yield and LAI)   included in the table 2 and 3. 
  
3. There are some errors in the citations,  so this should be checked and corrected (e.g. For example, Durbin and Willshaw [42] are quoted in the text, while in the list of references it is position – 43rd.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Dear Authors,
In my opinion, the explanations and the correction of the article are sufficient. Thank you for the opportunity to evaluate the article, and especially for the interesting methodological approach to yield evaluation. I wish you more successful works.