Determining the Capability of the Tree-Based Pipeline Optimization Tool (TPOT) in Mapping Parthenium Weed Using Multi-Date Sentinel-2 Image Data

Round 1

Reviewer 1 Report

Dear Editor,

This paper investigated the ability of the TPOT to map the Parthenium weed with Sentinel-2 multitemporal data in KwaZulu-Natal province, South Africa. The TPOT tool has been used for various research to find out the optimum classification and regression algorithms for the classification and prediction; however, this tool has not been widely used for remote sensing analysis. 

The authors have compared the results from the TPOT with a hybrid feature method (ReliefF, Support Vector Machines Backward, ExtraTrees) for mapping the Parthenium weed; finally, the authors realised TPOT computationally expensive for mapping the parthenium weed using multitemporal data.

The paper makes a practical contribution, but some aspects require further clarification.

Comments:

Line 28 - dimensional?

Line 91-100 - Please rephrase the paragraph; because it shows plagiarism from your previous research papers.

1) Feature Selection on Sentinel-2 Multispectral Imagery for Mapping a Landscape Infested by Parthenium Weed

2) Optimal window period for mapping Parthenium weed in South Africa, using high temporal resolution imagery and the ExtraTrees classifier

Line 101 - Please label the latitude and longitude on the map.

Line 156-157 - not very clear; please rephrase the sentences.

Line 156-170 - Is the Dark Object Subtraction method better than the Sen2cor atmospheric correction?

Line 159 - Which atmospheric bands were removed? Coastal aerosol, Water vapour and SWIR-Circus? Please mention the bands which you have used for the classification.

Line 160 - Which resampling method have you used?

Line 162-201 - Consider reducing the words and removing the formulas; because these are well-known algorithms. 

Line 213-231 - Please mention the code in the Appendix; it would be helpful for other researchers to map the parthenium weed and test this model.

Line 260-261 - 92.6% accuracy produced by TPOT or ReliefF-Svmb-EXT-TPOT? Vice versa?

Because in the abstract - you have mentioned, "The overall accuracies were 91,9% and 92,6% using the TPOT and ReliefF-Svmb-EXT-TPOT models, respectively." 

Line 289 - Table 4a and 4b are unclear; please mention the producer accuracy and user accuracy for each landuse and landcover class. 

Line 318 - Please label the latitude and longitude on the map.

Line 330-332 - This is unclear; please rephrase the sentences. 

Line 334-338 - Unclear; please rephrase.

Author Response

Dear Reviewer,

Kindly find "response to reviewers" file uploaded below.

Regards

Corresponding author

Author Response File: Author Response.docx

Reviewer 2 Report

Authors contributions:

According to the authors, the Tree-based Pipeline Optimization Tool (TPOT) outperform commonly used machine techniques, its capability to handle high dimensional datasets has not been investigated. Also, without the implementation of a robust classification algorithm or a feature selection tool, the large sets and the presence of redundant variables in multi-date images can impede accurate and efficient landscape classification of vegetation.

The main contributions in this paper can be summarized as: The TPOT can work well on a high dimensional dataset, such as the multi-date Sentinel-2 imagery, but at the expense of the computation cost; Combining a hybrid feature selection method with TPOT decreases the computational costs of the TPOT on a high dimensional dataset with slight increase in the classification accuracies.

According to the authors findings, the TPOT could perform well on data with large feature sets, but at the expense of the computational cost. The overall accuracies are 92-93% using the TPOT and other compared models, respectively.

The results from this study can be used in the management of invasive plants and their impacts in globally recognized biodiversity hotspots.

In their future work, the author will work on the development of algorithm that will be tested on larger feature sets, which will include a combination of vegetation indices, textures and spectral bands.

I have some reviewer notes:

At the end of the “Introduction” part, the authors can summarize their contributions. Not only what they will do in the paper.

“2.1. The Study area”. You can show a square of the study area with geographical coordinates. For example, in WGS84.

“2.2. Reference data”. It is not described, is 50 cm accuracy enough for your study? How the measurement accuracy affects your study? Also, you have to describe what criteria do you use to choose 70/30% sample sizes.

Line 156. You have to add information about QGIS – manufacturer and country of origin for this software system.

Line 174. “diff()”, sounds like “differentiate”. You can use “dist” or something that is appropriate to describe that mathematical function/operation.

Table 3. The classification accuracy is in %. The measurement unit is not visible. Also, check other tables for measurement units.

“Conclusions”. It is not clear how your results improve the known solutions in this study area.

Check your English language and stile. You can use online tools as Reverso, or something appropriate. (https://www.reverso.net/spell-checker/english-spelling-grammar).

I have some suggestions:

Improve the presentation of your results. Improve the presentation of used software and hardware tools. Make better summarization and comparison of your results with those from other authors. It will improve your contribution.

Author Response

Dear Reviewer,

Kindly find "response to reviewers" file uploaded below.

Regards

Corresponding author

Author Response File: Author Response.docx