Review Reports

Round 1

Reviewer 1 Report

The paper is interesting and thorough. It proposed a deep learning approach to identify floating vegetation suitable for snail habitat. The approach employs satellite imaging, but it is also validated with real-time images from drones. I have some small comments regarding the methods:

The image input stated in the manuscript is 256x256x8, however, in Figure 4, the U-Net architecture shows an input of 572x572x1. Were the images upscaled, or is the figure inaccurate with the U-Net model employed?

Overall, pixel accuracy is not suitable for unbalanced datasets; the authors state that only 4% is floating vegetation. Dice or Jaccard would be better.
In the discussion, the authors claim that: "This is not the concern for the imagery data used in this study, as the targeted 4 classes are not extremely imbalanced." which I honestly disagree with. However, floating vegetation segmentation achieved an excellent result, around 97% accuracy, thus the sentence could be changed to reflect that: This is not the concern for the imagery data used in this study, as the minor class achieved a very high segmentation performance around 97% accuracy.

The authors stated that: "Instead, the loss function was evaluated over the test data, and this served as an independent tuning metric of the performance of the U-Net." and "...segmentation accuracy of 94.5% for the test set and 82.7% for the hold-out validation/OOB set, over all four classes." I believe there has been some confusion regarding the validation and test set terminology.

The method section lacks details regarding hyper-parameter tuning, such as learning rate, batch size, and the number of epochs. The final learning rate was presented later in the manuscript. 

The authors stated that: "After training the U-Net model and fine-tuning the network parameters in an iterative process, the best-performing model...", what network parameters were fine-tuned? How was that process carried out? What was the selected hyper-parameter for the best-performing model?

Overall, the paper is solid, but some details that could improve reproducibility should be added.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Schistosomiasis is a major parasitic disease of poverty that affects more than 200 million people worldwide, mostly in sub-Saharan Africa, and is clearly linked to the construction of dams and water resource management infrastructure in tropical and subtropical regions. 
Changes in hydrology and salinity associated with the development of water management infrastructure can create favourable conditions for aquatic vegetation, which is a suitable environment for intermediate hosts of the gastropod parasite schistosome. 
Given the number - thousands - of small and large reservoirs, irrigation canals and dams that have been built or are under construction in Africa, it is essential to accurately assess the spatial distribution of risk environments that provide habitat for intermediate hosts of schistosomiasis from freshwater gastropods in rapidly changing ecosystems. 

However, standard snail monitoring techniques are laborious, time-consuming and provide information limited to small areas that can be manually sampled. In this study, the authors describe a novel framework for mapping the spatial distribution of suitable snail habitats at large spatial scales in the Senegal River Basin by integrating satellite data, low-cost, high-resolution drone imagery, and a computer vision technique using artificial intelligence (AI) called semantic segmentation. A deep learning model (U-Net) was created to automatically analyse high-resolution satellite imagery and produce segmentation maps of aquatic vegetation. 

The text is written based on scientifically accepted procedures and criteria. It introduces elements of novelty and uses examples to show the proposed solution and its validation.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf