Mapping Malaria Vector Habitats in West Africa: Drone Imagery and Deep Learning Analysis for Targeted Vector Surveillance

Round 1

Reviewer 1 Report

This manuscript used deep-learning and drone imagery to propose a method for identifying habitats for malaria vectors. I found the manuscript technically correct with detailed and clear methodology. However, my major critique is that, as the manuscript is currently written, the study is not well-justified by the biology of the system. I know that the study is in fact well-justified by the biology of the system but it I am familiar with Anopheles biology; readers who are not familiar with Anopheles and malaria biology may not be convinced that the effort involved in this manuscript is necessary. 

 

Specifically, I would like the authors to better describe the habitats used by larval Anopheles and why drone imagery is better than other EO datasets at detecting this habitat. Lines 29 - 32 is not sufficient to justify the development of the methods proposed in the manuscript. For instance, please describe the size and patchiness of water habitats that Anopheles larva is usually found in. Satellite imagery can detect water-bodies but not at the spatial or temporal resolution relevant to Anopheles breeding sites (< 1m^2), which clearly justifies why the proposed methods are superior to existing methods. Further, please further justify why the other land-use categories were selected in the context of Anopheles biology and why drone imagery is needed to classify them. Without this information, the land-classes seem quite general and the manuscript is more more about land-classification than detecting larval habitat. 

 

Otherwise, I have minor line edits:

Line 67-107: While I appreciate the thoroughness of the review of current methods, this section is highly technical and I think it would be best summarized in a table.

 

Line 129: main malaria vector?

 

Line 130-136: As the discussion later discusses the size and cost of the drones, I think it would be helpful for people unfamiliar with drones to put a picture of the different drones used in the supplement. 

 

Line 189: What is the resolution in meters?

 

Line 360: Missing a reference?

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Summary –

This study covers an important topic of mapping malaria vector habitat. It proposes a methodology to identify potential malaria vector habitat using drone imagery and deep learning techniques. Despite the limitations of high-resolution RGB drone imagery, the authors have applied this methodology to classify the drone images into a few land cover types including some that are relevant to malaria vector breeding habitats in the region with some success.

 

General Comments -

Although, this study applies recently developed tools to create a framework for this interesting and important application, it lacks clarity in some sections, lacks sufficient details in the methodology or justification of methodology in some parts. It also emphasizes the success of mapping specific land covers although not all land covers are representing vector habitat types. The methodology is useful to map some land cover classes, but it is important to clarify if those classes are potential vector habitat. 

 

This study needs to address some important points –

-       Do a better job explaining which habitats / land covers are relevant for the major malaria vectors in the study regions. The reasoning behind the choice of classes and connections to vector ecology needs to be stronger and made clearly.

-       Improve the training /ground truth data to better represent the land cover classes of interest. The ground truth masks (Figure 8) broadly cover the land cover types of interest. Isn’t it important to better delineate the boundaries of the land cover classes, especially if there can be confusion between classes? For this, it is important to define the different classes – e.g. Is emergent vegetation at the boundaries of vegetated water bodies part of the vegetated water body class? Should the ground truth polygons be edited to make corrections. 

-       Consider adding classes for the background areas that may conflict with the other classes of interest. 

-       It is important to address the seasonality of data collection and the analysis throughout the manuscript – from the introduction and methods section (description of the drone data collection) to the discussion – as that has implications for the results and the utility of this methodology for vector habitat mapping. Mention how and if the vector habitat changes by season. Lack of this information makes it difficult to judge the usefulness of study. 

 

Specific Comments –

 

Title – The title of the manuscript needs to be altered to specify the study maps malaria vector breeding habitat and not malaria vectors. 

Please change vector habitat to potential vector habitat throughout.  

 

Abstract – 

Please specify that the dice coefficient mentioned are for the best folds and not all results are above 0.68. Also clarify which land cover classes the dice coefficients are for.  Although some of the land cover classes selected are vector habitats, not all are. 

 

Introduction – 

Line 26 – With reference to “…classifying imagery into relevant habitat types.” – it is important to discuss that the EO data used are appropriate for mapping the habitat of interest.

Line 36 – can you explain why shape / configuration is important for epi studies?

Line 79 – please reword – this is unclear

Line 99 – This part is a little unclear. Is the German dataset a ground truth dataset with geo-referenced polygons and crop type labels? How was this different from the South Sudan / Ghana datasets – was it just a larger dataset? What geographic region does the German dataset cover? Were the different datasets covering different regions? Please explain why smallholder setting would be relevant. 

Lines 108-110 – Please mention which malaria vectors the study focuses on – In the Introduction you have mentioned 2 mosquito vectors in Africa but not mentioned the malaria vectors specific to West Africa that you are interested in. Similarly, please specify/ describe the breeding sites for these vectors and which land classes are associated with these. 

 

Methods – 

Using these deep learning techniques for identifying vector habitat from drone imagery is a great application. The land cover classes that are chosen, do not represent the entire landscape. Why were other classes not used to represent the background and train the algorithm? It seems the vegetation from these background areas (as seen in Fig. 5) might interfere with the vegetated classes you are interested in or the bareground might interfere with the tilled areas. Please explain how having large areas of bareground, scrub, other vegetation, etc. in the background is handled in your study and /or what implications that might have on the results. 

Please add details of computational needs or set up to the supplementary materials if the journal format allows that. 

 

Line 129 – Please correct from “main malaria in this …” to “main malaria vector in this …”

Lines 130-134 – How many drone surveys were conducted in each location? What is the temporal resolution of the imagery collected? Were surveys conducted in all seasons?

Line 136 – How were the drone images preprocessed? Please explain the steps in the manuscript or the supplemental materials. 

Figure 1 – Please label areas on the map. Are the 2 images showing the entire region where drone imagery was collected? If not, delineate on the map the region where drone imagery was collected. 

Lines 142-144 – Please explain the different types of water bodies that are considered habitat for Anopheles spp. – ponds, rivers, etc.? How do you define vegetated and non-vegetated water bodies? What kind of vegetation is present in the vegetated water bodies – is it submerged, emergent, floating? Are all the classes selected mutually exclusive?

Lines 142 – 148 – Please clarify which land cover types are relevant for Anopheles breeding habitat and which are not. 

Lines 153-155 – How many drone images were labeled?

Line 156 – What is the reason for two different tools? Was QGIS used because of the limited data allowance and internet issues for Groundwork? How were images chosen for the two tools?

Figure 5 – It will be useful to have the original drone image alongside the one with the overlaid polygons.

 

Results – 

Please add details about computation to the supplementary materials if allowed by the journal.

Can you discuss which classes had most confusion while mapping? For example - Based on the ground truth polygons and the difficulty in distinguishing between vegetated and non-vegetated water bodies, there will be some confusion between those 2 classes. Similarly, some confusion between land cover classes and the background should be expected. 

 

Line 260 – What are the yellow / green areas in the images shown in Fig. 7? It will be interesting to see what ground truth data were used for the images in Fig. 7.

Lines 298 – 300 – Why weren’t the ground truth masks improved after review of preliminary results? The ground truth masks broadly cover the land cover types of interest. 

Figure 8 – The polygons used for ground truth include a lot of the background area – especially for b and f. How does this impact the results? Can you show the prediction for all the classes in the image simultaneously? 

 

Discussion – 

Although, it is useful to note that such analysis can be used to identify areas for surveying vectors or for directing interventions, errors in boundaries of habitats will be a problem in identifying presence of a land cover at the edges – depending on how big the error is. 

Lines 311-313 – The classifier merely detecting the presence of specific land cover classes cannot be sufficient for a successful result. It is important that the classes of interest for vector breeding are being mapped correctly. Please specify which classes are mapped correctly and their relevance for vector habitat. 

Line 313-316 – Please reword. “…advertise the presence …” seems like an odd choice of words.

Lines 323-326 – Do you mean for 8F all pixels belonging to this class are mapped correctly?

Line 362 – Again, are classes corresponding to vector habitat being detected consistently? Please clarify. 

Lines 379 – 381 – Do you mean the land cover classes don’t represent all the possible habitats? Or that the selected areas for drone data collection do not include all possible habitats of the vectors? Discuss what seasonality of the data used in the study and provide context and explanation of what is missing and what is needed to improve the results. 

 

Conclusion –

Again, please specify which classes are important for vector breeding habitat. Roads and buildings are not important in that context and so having promising results for that class doesn’t prove the effectiveness of this method for your application. 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript “Mapping Malaria Vectors in West Africa: Drone Imagery and Deep-Learning Analysis for Targeted Vector Surveillance” is an important piece of work in field the land cover classification. The application of drone imagery and deep learning method can contribute greatly in the state of the art of optical remote sensing. The manuscript has been written well with excellent coherence and no flaws were found.  However, there is poor linkage with malaria vectors and the title is therefore not appropriate. It is merely a land cover classification and if title is revised on this way, I recommend this for the publication. However, if the authors want same title, I strongly suggest to use malaria vector data their relationship with landcover classification accordingly.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The title of this paper is very attractive and can arouse the interest of a large number of readers. But the problem is that the content is not well related to the title of the paper, or at least the authors did not clearly express it in the article. At present, the content presented in the article is the recognition of surface coverage types based on UAV images and depth learning methods. However, how these types relate to Malaria is the content that needs to be emphasized, but  unclear in the article. If the paper to be published under the current title, the authors need to strengthen the work and expression in this regard.

In addition, some specific suggestions:

1) Generally, the title of the form should be placed in the front of the form.

2) Some figures are redundant and have little information. It is recommended to delete them, such as Figure 1 and Figure 2.

3) The color matching and expression of the pictures in the paper need to be further optimized.

4) The abstract is somewhat lengthy and needs further organization and refinement.

5) The conclusion of the article needs to be further summarized and refined.

6) Whether it is possible to place an overall result of the study area?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 4 Report

The authors of this article have made targeted modifications or persuasive explanations to the issues that the reviewer was concerned about, and the quality of the article has significantly improved. It is recommended to accept.