Deep Learning-Based Automatic Extraction of Cyanobacterial Blooms from Sentinel-2 MSI Satellite Data

Round 1

Reviewer 1 Report

This is a good study, and I enjoy reading it. The paper can be published after addressing some issues and improving the clarity. Below are my suggestions:

1.      The abstract can be improved by adding more specific information/results; for example, what is the accuracy of DLN? What is the spatial/seasonal variability in the extracted algal blooms?

2.      Introduction: the motivation to use MSI imagery for monitoring HABs should be further improved. What are its advantages? Are there studies about the MSI in monitoring HABs in coastal/inland waters? Otherwise, DL some have some limits; you might provide some simple justifications.

3.      Figure 1. Information is not sufficient. The current figure is just a map. Suggest adding landuse/landcover around Lake Chaohu.

4.      Line 138: what is the usable image? What is the threshold of cloud coverage for filtering images? Are these images L1C/L2A?

5.      Table 1. Monthly distribution each year would be better.

6.      Section 3.2: the technical details should be supplemented. How to derive surface reflectance in Sen2Cor (e.g. settings)? What is the approach to resample to 10m resolution (tool?)? How to mask cloud and mask? The cloud masking is critical for extracting algae blooms, as clouds and floating scums have high signals. Have you considered the effects of land adjacency on monitoring algae blooms? The pixels around the land might be recognized as the algae bloom.

7.      Figure 8 can add some images of the difference between different method-derived HABs? Or adding the areas of algae blooms.

8.      Discussion: is it possible to extend DL models to OLI or other moderate-high spatial resolution instruments? An implication for the OLI-MSI constellation should make sense as MSI’s revisit time might not be sufficient to monitor variability in algae blooms. For example, Liu et al. 2021. Sentinel-2 and Landsat-8 Observations for Harmful Algae Blooms in a Small Eutrophic Lake; Li, J. A Global Analysis of Sentinel-2A, Sentinel-2B and Landsat-8 Data Revisit Intervals and Implications for Terrestrial Monitoring; Page, B. A harmonized image processing workflow using Sentinel-2/MSI and Landsat-8/OLI for mapping water clarity in optically variable lake systems.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This study is one of the first to undertake a deep learning-based approach to develop an automatic extraction model for CyanoHABs based on Sentinel-2 data (2006-2020) in Chaohu Lake, China. The CyanoHAB extraction model developed was trained with a CyanoHAB dataset generated by visual interpretation, and extraction results tested. The results were compared with three other automatic CyanoHAB extraction methods.

Major Comments:

1.       Abstract: The Abstract does not clearly summarize the goals of the study and methodology of the study. For example, the abstract states that high-precision and automatic extraction methods for CyanoHABs are lacking, but three “common automatic CyanoHAB extraction methods are tested.” It is unclear from the Abstract whether these common automatic extraction methods were tested against the developed DL method. Further information is needed on what the deep learning extraction model developed provides. There are a few instances where clarity could be improved by spelling out acronyms (i.e. U-Net, F1 in the abstract). High-level results should also be included.

2.       Methods: Was the CyanoHAB extraction model trained on a subset of the CyanoHAB Dataset, or on the entire dataset? This is not clear in the first paragraph in the Methods (it is detailed in Section 3.6.1). Including high-level information about the training and verification in the first paragraphs of the Methods would help clarify. In Section 3.6.3, it is unclear which dataset was used to evaluate the other extraction methods.

3.       Discussion: The verification of the applicability of the model to Taihu Lake seems to be a better fit for the Methods (and Results) sections. An interpretation of the results of the study and discussion of limitations should be included.

 Minor Comments:

Line 55: Please define FAI

Lines 428-430: This information detailing splitting of the training and verification datasets seems a better fit for the Methods section.

Lines 450-453: IOU is not previously defined. For all of these four results values, is the maximum (best performance) a 1?

Lines 472-473: It would be helpful to maintain a consistent order of the “other” methods. They are listed in the order of fixed, gradient, and Otsu in the Methods and Section 4.2.2. Gradient is shown before fixed in the images.

Lines 483-492: This might be a formatting issue, but I was unclear which text was paragraph text, and which table caption. The developed model is described as “DL model” but seems to be labeled “U-Net” in the table.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper statistically succeeded to give accurate output of CyanoHABs extraction by the DL model.  The model can be useful for any lakes although the ground truth should be acted. A process of application of the model is well described and discussion of sensitivity of the model to clouds is informative.   

Author Response

Thank you very much for your recognition of our study. The main purpose of this study was to prove that the DL-based method can automatically extract CyanoHABs in Chaohu Lake, as well for Taihu Lake. In the future, we plan to further train and test this method in more lakes and reservoirs.

Round 2

Reviewer 1 Report

Thanks for the revisions. The current responses satisfy me and can be accepted as is.

Author Response

Thank you very much for your approval of our revised manuscript.