Question about the data used in your papers

[SorourMo]

Hello, I would like to thank you for sharing your code on GitHub and congratulate you on your proposed models for cloud/shadow detection (Refined Unet v1 and v2). I have few questions about the dataset introduced in the refined Unet v1 paper (doi:10.3390/rs12122001): 1- Does all of the scene dates mentioned on page 9 (17 trainings, 8 validations, and 11 tests) belong to one location of path/row = 113/26?

2- For quantitative results reported in tables 1, 2, and 3, are the final refined cloud and shadow masks compared against masks extracted from each scene's QA band? If so, did you consider cirrus cloud pixels as a "cloud" class?

3- There are cases in QA band, where a QA value can be interpreted as both cloud shadow and cirrus cloud. For example, in the QA band, if a pixel has a value of 7072, it means that there is HIGH confidence of shadow and HIGH confidence of cirrus cloud for that pixel. That is the case with other values such as 7072, 7076, 7080, 7084, 7104, 7108, 7112, and 7116 (reference link:(https://www.usgs.gov/core-science-systems/nli/landsat/landsat-collection-1-level-1-quality-assessment-band?qt-science_support_page_related_con=0#qt-science_support_page_related_con))

It would be great if you could share the values of QA band you used for both classes of cloud and shadow in your experiments. If not possible, answering these questions might help researchers to figure out this issue themselves.

Thank you!

[92xianshen]

Dear @SorourMo ,

Thank you for your kind suggestions. I will deal with them ASAP.

Libin

[92xianshen]

Dear @SorourMo ,

Thank you for your comments, and here are my responses.

1- Does all of the scene dates mentioned on page 9 (17 trainings, 8 validations, and 11 tests) belong to one location of path/row = 113/26?

A: Yes, we just make use of the scenes of the same location: Landsat 8 OLI, path 113, row 26.

2- For quantitative results reported in tables 1, 2, and 3, are the final refined cloud and shadow masks compared against masks extracted from each scene's QA band? If so, did you consider cirrus cloud pixels as a "cloud" class?

A: Thank you for your comments. We use the QA band as the label in the training and reference in the test. We have checked the code labeling pixels of clouds and shadows: pixels of 480 and 992 are labeled as clouds, and pixels of 328, 392, 840, 904, and 1350 are labeled as shadows. These values are presented in Table 7-3 in the pdf file of LANDSAT 8 SURFACE REFLECTANCE CODE (LASRC) PRODUCT GUIDE.

In our paper, the quantitative indicators (P., R., F1, and acc.) are computed between Refined UNet and QA generated above, which are able to indicate the specificity and sensitivity of the model from the global perspective (on the entire test set). But we prefer to show and evaluate the results visually because we focus more on edge-precise segmentation, which is a low-level visual task. It can be seen that Refined UNets (v1 and v2, we will present v3 later) have achieved great segmentation improvement in the areas of cloud and shadow edges. This is why we evaluate quantitatively and qualitatively.

3- There are cases in QA band, where a QA value can be interpreted as both cloud shadow and cirrus cloud. For example, in the QA band, if a pixel has a value of 7072, it means that there is HIGH confidence of shadow and HIGH confidence of cirrus cloud for that pixel. That is the case with other values such as 7072, 7076, 7080, 7084, 7104, 7108, 7112, and 7116 (reference link:(https://www.usgs.gov/core-science-systems/nli/landsat/landsat-collection-1-level-1-quality-assessment-band?qt-science_support_page_related_con=0#qt-science_support_page_related_con))

It would be great if you could share the values of QA band you used for both classes of cloud and shadow in your experiments. If not possible, answering these questions might help researchers to figure out this issue themselves.

A: As presented above, we used the Level-2 Pixel QA band to label cloud and shadow pixels. It is much appreciated that your comments can help us improve the quality of our work.

We hope our explanations can be accepted, and once again, thank you very much for your comments.

Libin