Hyperspectral Feature Selection for SOM Prediction Using Deep Reinforcement Learning and Multiple Subset Evaluation Strategies
Round 1
Reviewer 1 Report
The authors propose a hyperspectral feature selection method based on deep reinforcement learning, which using multiple subset evaluation strategies, named RLFSR_net and RLFSR_cv. To test the performance of the proposed method, inverse model is established on airborne hyperspectral images, which can achieve good inversion results. However, there are still many problems that need to be addressed.
1. In Section 2.2, the hyperspectral regression deep neural network is trained for the feature evaluation, please explain in detail the specific process.
2. In Section 3.1.2, 90 soil samples were processed by several feature extraction methods, what is the purpose of this step and why these features extraction methods are taken?
3. In line 397-399, the article describes the processing of soil samples, please add a more detailed description and add a reference for the SOM measurement method, e.g. New methods for improving the remote sensing estimation of soil organic matter content (SOMC) in the Ebinur Lake Wetland National Nature Reserve (ELWNNR) in northwest China.
4. There are several inconsistent descriptions of Pearson’s correlation coefficient in the article (in line 101 and 522), please modify to ‘Pearson’s correlation coefficient’ and add reference.
5. Some state-of-the-art hyperspectral feature selection methods are suggested to be added for the comparison analysis.
6. The color scheme of the picture varies widely, please check it.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 2 Report
In this paper a feature selection framework: reinforcement learning for feature selection in hyperspectral regression, is proposed.
The paper is well presented and structured. However, not only quality metrics but also computational times should be considered in the algorithms assessment.
The level of the use of the English language is unsatisfactory to be included in this publication. Authors whose primary language is not English are advised to seek help in the preparation of the paper.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
This work describes a method to select hyperspectral features to perform soil organic matter predictions. The main advantage is the introduction of deep reinforcement learning for feature selection. The results are later evaluated using different strategies and it shows how the proposed method provides better performance results using less features. I find the text is well written and the methods as well as the results are clearly presented. This works shows how the use of re-inforced learning for feature selection can improve the results and provides as well insights on how features are selected by different methods. I recommend the publication of this work. I have only a few minor comments for the authors to consider:
- I think in the text below equation 8 there is a confusion between "\gamma" and "r". My undestanding is that "\gamma" is the discount factor as introduced in line 241. However in lines 320-324, it looks like "\gamma" and "r" are exchanged. I suggest to check the lines 320-324 again.
- In table 2, the results of CARS are the closest ones to FLFSR results. However, there is a big difference on the performance of PLS which is very bad for CARS while not so much worse for the FLFSR methods. Why is CARS so much worse with PLS? Is it due to the numbe rof features which is significantly higher?
- In Figure 7 there is a line labelled as "ALL". Is it an average? How is it defined? I think the text does not describe the meaning of that line.
- Table caption in line 541 is not correct. The table number shall be 3, I think and the text looks like it was forgotten to update.
- The SOM prediction maps in Figure 1 and Figure 6 have low resolution and it can be difficult to see details on them.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
Once the considered aspects have been addressed the paper can be accepted.