Improved Fusion of Spatial Information into Hyperspectral Classification through the Aggregation of Constrained Segment Trees: Segment Forest

Round 1

Reviewer 1 Report

 

This paper presents a method for hyperspectral imaging classification. The PCA is applied before using SVM to classify pixels. The novelty of this paper is limited since the authors propose an optimzation process to obtain spectral information. Here are some points need to be clarified:

- Different band selection methods should be included in literature review. For example, you can mention this paper, 
DOI: 10.1007/978-3-030-68790-8_25.

- Please give a reason for choosing 3 datasets for experiment. Other benchmark dataset such as HyTexiLa can be considered. 
- All the obtained results in this paper are not compared with those in the state-of-the-art. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript outlines a method (segment-forest) for the classification of hyperspectral images based on combining spatial and spectral information in the image pixels through using segment trees. The manuscript also presents the results on several datasets in comparison to several spatial-spectral classification methods.

The introduction of the manuscript nicely outlines the types of spatial-spectral learning as feature-level fusion, decision-level fusion and neural network based deep learning. The SF method does not use neural networks and is therefore relatively parameter sparse. The SF method seems to build up on previous work by the authors (segment-trees) and is apparently an adapation of the method, such that a segment-forest is (to my understanding) a set of segment-trees that span an image. While the vocabulary used is evocative, I did not see any place where it is clearly defined what a segment-forest is and how it differs from a segment-tree. I think the work should be presented with enough detail/clarity that the relationship to segment-tree should be evident to the reader. Also, how does one select/choose/assess how many segment-trees should be generated in a segment-forest?

While a basic exploration of the effect of the introduced parameters is given in section 3.1 (not mentioned in abstract), there is no guidance provided on how one must choose these parameters. What do they represent? What is the basis for choosing a value -- perhaps something related to the number of classes or some distribution charactersitic of the pixels in the image?

L144-146: What is step 3 of the process? What does "construct the segment-forest" mean? The process should explain it in some detail rather than saying that "was extracted to build the segment-forest".

Overall I feel the title is not indicative of the work presented. The "novelty" is in the adaptation of how the spanning tree is constructed on the principal component of the image, and "based on segment-forest" provides the suggestion that segment-forest is an older method. While that misunderstanding is clarified on reading the abstract, the title still has provided no information on what the presented paper is about. I suggest the author's reconsider a more informative title to the work presented like "Improved fusion of spatial information into hyperspectral classification through aggregation of constrained segment trees" or something of that nature. Perhaps even indicating that this is an adaptation of segment-trees, and the main focus is on 'spatial filtering' of classification results obtained through another method.

The captions for Figure 7 and Figure 8 are very unsatisfactory, and do not help the reader in any way to understand what the authors are trying to demonstrate. Without a detailed caption stating not just what is shown, but what is to be noticed, this remains a mosaic of colorful images and the panel labels (a, b, c) just point to other acronyms that one has to look up in the abstract. What is the red rectangle in figure 7? This should be in the caption.

The entire section 3.3 where inter-method comparisons are created does not report observable facts in the "results", i.e. the figures. Instead it turns out to be a series of claims, more like discussions, without any citations or support for the claims. Such treatment of comparison with other methods is not convincing, and does not fit under results. The provided discussion points indicates that the authors have compiled a valuable set of information on the run-time or implementation difficulties of the various methods. If this is the case, I urge them to develop a figure showcasing the relative computational effort-vs-reward relationship of the different methods. If not, I suggest to limit the claims to those that can be supported independently.

Another overall issue I have with the presentation of the comparison is that it is compared with a variety of different spatial-spectral methods. While I appreciate the attempt to be comprehensive, the real meaningful comparison is between segment-trees and segment-forest, since the currently presented method is essentially an adaptation of the segment-trees method of filtering initial classification results. A richer explanation of the pros and cons between the segment-trees and segment-forest would help the reader understand the achieved improvement. Comparing it to SVM (like in L357-360) is not a meaningful or honest comparison point.

L185-187: What are the height and width of the first prinicipal components? This is not terminology familiar to me, and the given explanation does not quite indicate how it functions.

L316-317: More explanation is needed for the claim maded that "is challenging to obtain and so to relationship is set as a constant. What is it challenging in comparison to? The method has been tested several times. Also, is the constant relationship between pixels set as a constant by the authors? If not, what study should be cited here to provide context/evidence?

L 331: "Compared with ST, the segmentation between different similar regions in SF is disconnected, so better classification results are obtained in filtering." Please explain better what this means that "different similar regions are disconnected in their segmentation". ANd why would this result in better classification results?

Figure 2: Please show the ground truth image so that the context of the results are included

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

In this paper, the authors proposed a hyperspectral image (HSI) classification method based on segment-forest(SF). The aim of this approach is to improve the HSI classification by considering the spatial and spectral features in the classification task. The proposed model seeks firstly to reduce the high dimensionality of HSI using the principal component analysis (PCA). Moreover, the support vector machines (SVM) has been applied to perform the initial classification results. Finally, the segment-forest has been developed in order to optimize the initial classification results and get the final classification results.
For experimental results, the proposed approach has been validated on three public HSI datasets, including Salinas, WHU-Hi-HongHu, and XiongAn.
The proposed idea is interesting, however, some revisions have to be made and some parts of these experiments are not complete to claim the advantage of the proposed method :

1) The contributions of this paper seem to be not clear to the reviewer. Please clarify the differences between the proposed method and some existing hyperspectral image classification. For example, which contributions are existing and which ones are your own?

2) In experimental results, did the authors randomly select the training samples? If that is the case, what happens when you change the selected pixels (another random sampling of training data)?

3) The authors claimed that the proposed method outperforms the state-of-the-art classification methods of HSI. How do they explain the high performance of the proposed method?

4) Could the authors explain what is the motivation to use PCA rather than other dimensionality reduction methods, e.g. autoencoder, band selection, etc. 

5) I suggest the authors to add these recent references  in the manuscript which are related to HSI classification using spectral and spatial features :

- Deep neural networks-based relevant latent representation learning for hyperspectral image classification, Pattern Recognition, 2022.
- A shallow network for hyperspectral image classification using an autoencoder with convolutional neural network, Multimedias tools and applications, 2021.

6) The English and format of this manuscript should be checked very carefully

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

The authors have revised this manuscript carefully according to my questions. I have no further questions about this manuscript. It could be accepted.