Unsupervised Characterization of Water Composition with UAV-Based Hyperspectral Imaging and Generative Topographic Mapping
, Adam Aker, Lakitha O. H. Wijeratne, Shawhin Talebi, Ashen Fernando, Prabuddha M. H. Dewage, Mazhar Iqbal, Matthew Lary, David Schaefer, Gokul Balagopal and David J. Lary *
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors have presented a very interesting paper in which they used a hyperspectral camera with a drone to gather data on a lake and use these data to classify different types of surfaces. The topic is updated and linked to the journal's scope. Nevertheless, there are some aspects of the paper, mainly linked to the labelled surfaces and the training dataset, which must be improved. Following, I include a series of comments aimed at enhancing the quality of the paper:
1. It is not recommended to use acronyms in the abstract; I suggest deleting the acronym of UAVs.
2. In the abstract, the introduction of the problems is too extensive. The authors should reduce the problem statement to just two lines.
3. In the abstract, the authors have to highlight their results by adding the numerical values that summarise the proposed system's performance.
4. In the introduction, to provide a context on the use of characterization of water composition, the authors should include the proximal sensing based on sensors as described in the following recent papers:
a. (2024, February). Optical Sensor for Water Bacteria Detection using Machine Learning. In 2024 11th International Conference on Computing for Sustainable Global Development (INDIACom) (pp. 603-608). IEEE.
b. (2024). Combination of Machine Learning and RGB Sensors to Quantify and Classify Water Turbidity. Chemosensors, 12(3), 34.
5. The structure of the paper must be added at the end of the introduction.
6. Figures must appear after being cited in the text. Please change the position of Figure 1 to another Figure with the same problem.
7. The authors have to reconsider the label of "Grass" in Figure 1. The spectral signature does not correspond to the typical signature of grass. It is not possible to see the chlorophyll peaks as expected in the signature. Is this grass in bad health, or does this point correspond to the ground? Please check and modify the label. Check it also in line 190 and in other parts of the text and figures. There is another problem with the "Algae" label. The authors have labeled as algae a particular type of algae living on the shore. Nevertheless, algae also represent the microorganisms in the phytoplankton. Thus, an alternative name should be used. Moreover, it is recommended in subfigure d that the general map of the USA be added.
8. In section 2, please add the manufacturer name and location for the drone, camera, spectrometer, acquisition computer, etc.
9. Check this sentence: "…(a) of Fig 7 where the log10-reflectance values…". There might be an error with the number of figures.
10. The authors have to add the reference for the NDWI.
11. The authors have to provide the details of the used dataset to train the GTM model. They should include the number of samples and pixels for each type of surface. This information must be provided in the methodology. If possible, consider distributing the data in one of the images. This is very important given the labels used since grass might represent the ground, and algae represent only algae inshore, not the general algae.
12. The results need a validation stage. Showing the general classification of the image and comparing it with an image classified by the authors manually can serve. This data can also be summarized as a confusion matrix commonly used in classification problems.
13. In the discussion, the authors have to include new content. They should compare their results with existing research on this topic, evaluate their proposal's performance, and summarize this information as a comparative table. Moreover, it is recommended that the section be divided into different subsections to facilitate their comprehension.
14. Future work has to be included at the end of the conclusions.
Author Response
We sincerely appreciate the time and effort that you dedicated to reviewing our manuscript and are grateful for your insightful comments and suggested improvements. Thank you!
Please find a PDF of or responses to your points. We appreciate you!
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis is a very well written paper that is based on a thorough body of work by the authors. In their study, they attempted to demonstrate how the Generative Topographic Mapping technique could be used to visualize high-dimensional hyperspectral imagery and extract spectral signatures corresponding to unique endmembers present in the water. To demonstrate the effectiveness of the approach they collected data across a North Texas pond using a small UAV equipped with both a hyperspectral sensor and an upward looking spectral irradiance meter for absolute water reflectance calibration. As a further validation they show that using the retrieved end members they were able to map near-shore algae and a localized contamination source simulated by a rhodamine tracer dye release. Both the algorithm and the experiment are comprehensively described. Small UAV’s are currently more and more used to map inland water bodies and check for the presence of contaminants. The technique the authors present and validate has two very significant advantages over more standard approaches. It can detect the presence of unexpected contaminants since it does not a priori require a full spectral signature library but instead uses only the data to establish the end members. Since their technique can also handle non-linear signature mixing it is particularly well suited for the turbid water environment with vegetation present typical of inland waters. The sources of the non-linear mixing in that latter case are the diffusion of the water return due to scattering by waterborne particles which mixes signatures from out of pixel sources. The effects are generally referred to as adjacency effects. Vegetation which is very often partly transparent in several spectral intervals is subject to a layer thickness effect which also modifies the signatures non-linearly (exponentially) and adds to the per pixel cover fraction. The results have significant potential impact in improving the retrieval of remote sensing over complex ecologically significant areas. For these reasons I strongly support publication of this paper as it should be extremely useful to the part of the limnology community using hyperspectral remote sensing as a tool.
I have listed below some suggestions for clarifications which would in my opinion make the paper more readable in particular to the Oceanographic and Limnological communities.
Caption of figure2 “M-many RBF”, RBF needs to be explicitly defined. It is defined after the figure and this point should be indicated in the caption or the figure should be moved after the RBF definition.
Line 132 “responsibility” Please explain the concept of responsibility? Please describe the analogy on which the name responsibility was chosen. I know you have clearly defined it mathematically but responsibility is a word which carries a great deal of significance in common use and some justification for its use in this technical context should be explicitly given if possible. As an example, in his original paper Shannon quite brilliantly explained his use of the logarithm probability of a surprise as relatable to being given more information.
Line 146, Quadratic Discriminant Analysis QDA (needs reference and brief description of how it functions.
Line 162 It is the upwelling radiance which is assumed to be Lambertian due scattering and no the water surface itself which in a pond is much more specular than diffuse particularly under low wind conditions.
Line 263 Why NS3 value of 0.4725 chosen. Seems awfully specific. Please explain the rationale.
BTW The suggestion of generating an autonomous boat optimum path for ground truth collection is brilliant.
Below are some minor typos I noted.
Line 33 “on-board compute” should be replaced by “on-board computing”.
Line 94 “HSI were collected” should be replaced by “HIS was collected”
Line 98 “processes” should be replaced by “process”
Line 98 “pseuedo” should be replaced by “pseudo”
Author Response
We sincerely appreciate the time and effort that you dedicated to reviewing our manuscript and are grateful for your insightful comments and suggested improvements. Thank you!
Please find a PDF of or responses to your points. We appreciate you!
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsGeneral comment
It has been presented in the manuscript that a Bayesian realization of the Self-organizing Map GTM was used to visualize high-dimensional hyperspectral imagery and extract spectral signatures corresponding to unique endmembers present in the water. The GTM was applied to extract endmembers and to efficiently map the abundance of near shore algae. Although the manuscript holds a merit to be published, some efforts are needed to be fixed as the details of methods and results should be more detailed. I will address myself in detail as follows.
Specific Comments
1. It is recommended to change the color of the sample points in Figure 1b to a high-contrast one, which is unclear in the current form.
2. Some content in part 2.1 is not cited. It is recommended to review it and add necessary citations.
3. A legend should be added to Figure 4. It is currently difficult to understand the content of Figure 4.
4. From the context description, the HSI reflectance obtained by this method is not corrected for skylight. Will this affect the results?
5. Which part of the water spectrum does GTM obtain?
6. How is the endmember defined in this study? How to determine the endmember spectrum in this study?
7. Line 263: How is the threshold determined?
8. Much content in the results section can be transferred to the methods section. The results section only retains the experimental results and reduces the description of the methods.
Author Response
We sincerely appreciate the time and effort that you dedicated to reviewing our manuscript and are grateful for your insightful comments and suggested improvements. Thank you!
Please find a PDF of or responses to your points. We appreciate you!
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsReview notes may be find in the attached file.
Comments for author File: Comments.pdf
Author Response
We sincerely appreciate the time and effort that you dedicated to reviewing our manuscript and are grateful for your insightful comments and suggested improvements. Thank you!
Please find a PDF of or responses to your points. We appreciate you!
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsI have no additional comments.
Reviewer 3 Report
Comments and Suggestions for AuthorsMost of my concerns were addressed.
