Feature Selection on Sentinel-2 Multispectral Imagery for Mapping a Landscape Infested by Parthenium Weed
Round 1
Reviewer 1 Report
The manuscript presents a comparison of feature selection approaches, in order to map Parthenium weed using Random Forest classification on Sentinel-2 images. The comparison and analysis of the results are well described. But as the aim of the comparison is to identify the optimal feature selection approach in terms of accuracy and computational complexity, I would like also to see the computational time needed for their implementation.
Below you can find my remarks:
2. Material and Method
2.2 Reference data
Line 118: How did you select these three ratios? In [30] they are dealing with SVM and feature selection, but I could not find in any suggestion for selecting these ratios.
3. Results
3.1 Comparison among investigated features algorithms
In Figure 2, better to put ‘Number of features’ instead of ‘Number of bands’ in x-axis name. You don’t have only bands as features, but also indices, actually mainly indices. I would suggest also to mention again in this paragraph what the features are (75 VIs + 10 S2 bands). In the legend of Figure 2 also better to put what 1,2,3 stands for, meaning to put 1:3,1:1,3:1. In addition in the caption of Figure 2 (line 256) you wrote ‘Optimal subsets in brackets’, but I could not see them in the Figure. You would be better write also ‘Optimal number of features’ or ‘Optimal subset of features’.
3.2 Comparison of performance between peak accuracy …..
3.2.1 1st training set
I would like to see the number of features as an extra column and not in parentheses in Tables 2,3,4.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 2 Report
It is globally important to accurately classify the area infested by weeds as these weeds could possibly have a huge impact on the environment and agriculture. In the present study, the authors explore the efficacy of different feature selection methods belonging to 5 different groups and set to find out which method is the best for reducing the dimension of sentinel 2 bands and vegetation indices and gives the best accuracy for classification of the area infested by parthenium weed. I only have a few minor concerns about the manuscript:
LL 8- 10- “Hence…….” As of now, there is no proper flow of logic from the previous sentences to this sentence. There should be a connecting sentence to explain this properly.
Abstract: the results mentioned in the abstract is unclear. Rewrite to improve the comprehensibility.
LL 48- 50 – The flow of logic is not proper. There should be a connecting sentence here.
Figure 1- You should mention which bands constitute the RGB image on the left panel. Also, mention the wavelengths of bands used.
LL-204-206- logic is not clear
LL 235- Why the number of trees was kept at 10 for every RF run. Wouldn’t different methods differ in their performance based on the number of trees? Should explain
Results-It might be good to give classification accuracies of other classes and kappa coefficient of the whole effort as well (maybe as supplementary information).
Fig 3- Check figure legend and correct it.
Figure 3 and 4 almost represents the same information. It may be better to keep only figure 3???
Author Response
Please see the attachment
Author Response File: Author Response.docx