Parameterization of a Bayesian Normalized Difference Water Index for Surface Water Detection

Round 1

Reviewer 1 Report

this manuscript is a good presentation of a new bayesian based approach to calibrating an NDWI index for detection of water bodies, in particular in areas where small lakes are present. The interest of NDWI based detection of water bodies is very significant, in many areas such as migration of coastlines, flooding of extensive areas, tidal nuisance, among others. This manuscript is therefore very useful in that it presents a new way of calibrating NDWI indices against Sentinel-2 data and to optimize such indices. The methodology is sound, well explained, without redundancies, which is an accomplishment given how difficult the concepts of bayesian inference can be to non-users of uncertainty quantification concepts. I found the manuscript sound, novel in the use of this bayesian inference approach, and thorough in its assessment of the utility of the new NDWI index, and in how the threshold question needs to be approached. I therefore recommend this manuscript for publication, with some minor modifications.

First, the question of the Kappa parameter is for experts. It could be introduced better. I myself did not know anything about it, and I'm not sure I know more now than before. Citations could be improved for the introduction of NDWI, and I seem to detect a certain bias towards specific authors. This is a very large field, and I'm sure there is other citations for NDWI. Please include the latest Dai et al, 2019 paper on NDWI for coastal zones, it is highly relevant here. Finally, I feel like the discussion of the results could be shortened without losing too much. I found myself confused by too many interpretations or explanations towards the end of the manuscript. I think the method speaks for itself, and I don't believe readers will be overly interested in minute details of how much exactly the fit occurs or not for specific small bodies of water. A lot of these details could potentially be offloaded to an annex. 

Thank you for considering me for this review, I very much appreciated the manuscript, and hope you will publish it.

Best regards,

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Please see the attachment. 

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Geosciences-822792

Overall:

Authors have proposed another modified index to map water bodies using 4 bands (10m) of Sentinel 2b images acquired from 2015-2019. This index combines the TOA values from 4 bands after estimating 8 parameters.  While they have explained the mathematics behind how these parameters were estimated using Bayesian methods, there is no explanation of their physical meaning.

In the numerator of the proposed model (equation 2), values of blue, green and red are added together after multiplying the respective parameters, and contrasted against the NIR values (band 8).  On the other hand, the denominator values of G, R, and NIR and combined after multiplying the respective parameters, and contrasted against blue band (2). In remote sensing terms what does this mean?  I agree this combination of spectral values and parameters worked to identify water bodies, but authors have to explain the remote sensing basis of this equation.

Will this set of parameters work universally or they should be computed for each study?  This needs to be explained clearly in the abstract, results, and conclusion sections. Authors are recommending that this index can be used for delineating flooded areas.  Can the proposed equation used as such or the parameters have to be estimated every time?  This is a major limitation in several remote sensing studies.  First study will propose an index with parameters, and future studies will use them as such.

I agree that the proposed index does not initial calibration of the threshold value, whereas the original NDWI will require selection of an optimal threshold value which will require time.  However if future studies have to calculate the parameters for the proposed index it will require additional time.

Authors have computed the accuracy of the classified images using reference data generated from orthophotos.  What was the acquisition date for the orthophotos?  Surface areas of the water bodies change frequently.  Authors used Sentinel 2b images acquired from 2015 – 2019.  How did they account for the temporal differences in the surface area in these images while comparing them to the orthophotos?

 

 

Abstract: There is already a modified NDWI index, which is used to overcome the effect of built-up land. How is the proposed modified index, varies from MNDWI?  Mention this briefly in the abstract.

Introduction:

Authors have mentioned the opportunities and limitations of various normalized and other indices for mapping water bodies. Then (in line 77), they abruptly mention that they are proposing another modified version of NDWI.  They have to make a stronger case as to why we need another modified index.

Material and Methods:

Following equation 2, authors mention bands 2, 3, 4 and 8.  While bands 3 and 8 are mentioned earlier (line 95), the other two mentioned in Table 1 (page 6) which appears much later.

Line 149.  ‘Acquired in 2015-2019’ will be easier to understand instead of ‘referred to the 2015-2019.

Line 106: What will be the range of NDWI(m) values?  Later authors set of a threshold value of > or = 1 for identifying water bodies (Line 157).

Line 150-152. Here and also in Figure 1, authors mention the “raster of actual water reservoirs obtained from the orthophotograph of the study area”. Since the surface area of the reservoir is dynamic how did they account for the differences?  This must be explained in detailed here.  Already mentioned in the overall comments.

Line 153: Authors mentioned for each parameters, a prior distribution was selected.  How was this done?  What are the range of values for each parameter?  Are they directly or indirectly related to each other?  More information is needed here.

Line 157-158: How did they account for temporal differences between the digitized orthophoto map and satellite images (acquired from 2015 – 2019)?

Results and Discussion:

Lines 212 – 227. Should these lines be under methods section?

Lines 270 – 272. Should the sensitivity analysis be mentioned in the methods section?

Lines 388 – 406: These lines must be under methods.

Line 407: What about negative agreements and what will be those kappa values?

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

I am satisfied with the revised manuscript, and I have no further comments and suggestions. 

Author Response

We thank the reviewer

Reviewer 3 Report

In their response, authors have placed the emphasis on the outcome of Bayesian methods for estimating the 8 parameters. Authors must provide a physical meaning for the resultant model (equation 6 in page 8) with 4 spectral bands and 8 parameters. In the numerator this model is summing the values in the visible bands (2, 3 and 4) after assigning individual weights to them. Then the NIR values (band 8) is subtracted from it. In other words, it is a contrast between the visible and NIR band values, but values in some visible bands are given twice the weight. In the denominator, values of green, red and NIR are summed, and the fraction of blue value is subtracted from it. There is a physical explanation for these weights, and addition/subtraction which must be explained.

Authors have checked with the authorities to confirm that no new reservoirs were built, or existing ones were demolished between 2015-2019. However, water surface area in a reservoir can change frequently. Following precipitation events, water surface area can be high and can be low at other times. Since this study focused on small reservoirs (0.1 ha – line 83), the amount of water could potentially change from season to season. In extreme cases, the water body could be entirely dry at some point in time. The image will report those pixels as bare ground, whereas the validation data from the orthophoto would say it is a water body. There are chances for the opposite situation to happen. For example, large precipitation could result in the filling of low-lying areas which would show up as water bodies in the satellite image. However, those areas would be non-water in the orthophoto-derived validation set. How did the authors account for these differences since they do not have high resolution orthophoto for each study year?

Authors mention that based on the overall accuracy values (99.7% and 99.71%) and kappa agreement indices (0.825 and 0.837), the performances of both existing and proposed indices are quite similar (lines 492-493). Further they state the index lacks precision in the detection of very small reservoirs between 20 and 500 square meters. I agree that NDWI (old index) requires selection of optimal threshold which requires time. However, estimating the parameters with Bayesian techniques also requires time since new studies have to (re)estimate these parameters. Is all the extra work associated with Bayesian modeling for estimating the parameters worth for this incremental increase?
Authors must compare the results from the proposed index to the corresponding ones from MNDWI.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 3

Reviewer 3 Report

Thanks for addressing my comments about the physical meaning of the proposed index.

1. Temporal differences: In their response document authors state that “due to their geometric characteristics, the reservoirs in the area of study are rarely completely dry”. Hence, they are confident that the high-resolution imagery acquired in 11 June 2018 (Google Earth), represents the status of the water bodies for all years (2015-19).

I recommend that authors stress this in the abstract and conclusions as well, so future studies (or researchers) are aware of this especially when applying this technique in areas where the surface areas of waterbodies shrink and expand on seasonal and annual basis.

2. Newly added Figure 11 is helpful to understand better precision. However, its caption has to include sufficient detail to explain that the proposed index is better identifying more pixels in comparison to NDWI.

3. Optimal parameters: In this (R2) version, authors have added new text that states that “the optimal set of parameters does not need to be updated when new spectral band images are available…”. This statement needs further clarification, because in the previous (R1) version, they added new text that stated “To ensure an optimal performance of the proposed index, the application of the Bayesian parameterization is suggested for each case of study to be investigated” (lines 273-277).

Once again, I recommend that authors further clarify these statements. Are they implying that when new spectral images are available for this study area, the parameters do not have to be updated?  Or is it applicable universally?  Also does this index with these parameters work only for Sentinel-2b data? 

 

After this article is published, future studies will take the proposed index and will apply it to other regions using Sentinel-2b or other satellite data with these bands.  These studies may not have access to high resolution image for verifying their output.  It is imperative for these authors (and anyone) who propose a new index to stress the assumptions/conditions under which it will work.

Author Response

Responses to Reviewer #3

Thanks for addressing my comments about the physical meaning of the proposed index.

1. Temporal differences: In their response document authors state that “due to their geometric characteristics, the reservoirs in the area of study are rarely completely dry”. Hence, they are confident that the high-resolution imagery acquired in 11 June 2018 (Google Earth), represents the status of the water bodies for all years (2015-19).

I recommend that authors stress this in the abstract and conclusions as well, so future studies (or researchers) are aware of this especially when applying this technique in areas where the surface areas of waterbodies shrink and expand on seasonal and annual basis.

Action taken: The following sentence has been added in the Abstract (page 1, lines 19-21): “To assess the effectiveness of the index, a reference image, representing the actual reservoirs in the study area, was used”.

The following sentence has been added in the Conclusion section (page 15, lines 518-524): “In this study, a reference image of actual reservoirs, obtained by manually digitalizing the reservoirs visible in high spatial resolution Google Earth images of the study area. Reservoirs in the area of study are affected by slight seasonal variations, therefore the reference map can be considered quite representative of the actual reservoirs in the catchment. It has to be remarked that, to ensure a reliable performance of the proposed procedure, the use of an accurate reference image for testing the effectiveness of the index is highly recommended when the reservoirs in the area of interest are subjected to seasonal or annual variations”.

 2. Newly added Figure 11 is helpful to understand better precision. However, its caption has to include sufficient detail to explain that the proposed index is better identifying more pixels in comparison to NDWI.

Action taken: The caption has been modified as follows (page 12, lines 409-410): “Details of the reservoir maps in Figure 10. For the reservoir at the bottom of the images, the NDWI_m is more effective in detecting the water cells along the boundary of the water body”.

3. Optimal parameters: In this (R2) version, authors have added new text that states that “the optimal set of parameters does not need to be updated when new spectral band images are available…”. This statement needs further clarification, because in the previous (R1) version, they added new text that stated “To ensure an optimal performance of the proposed index, the application of the Bayesian parameterization is suggested for each case of study to be investigated” (lines 273-277).

Once again, I recommend that authors further clarify these statements. Are they implying that when new spectral images are available for this study area, the parameters do not have to be updated?  Or is it applicable universally?  Also does this index with these parameters work only for Sentinel-2b data? 

In the reviewed manuscript, we specified that the parameters have not to be updated for the analysed case of study but the application of the procedure to a different area requires to perform the Bayesian parameterization of the proposed index, as well as the use of different satellite images.

Action taken: The following sentence has been modified (page 14, lines 468-469): “Moreover, for the analysed case of study the optimal set of parameters does not need to be updated”.

The following sentence has been added (page 14, lines 472-476): “It has to be remarked that the optimal parameter set here assessed is not valid for any area of study. Thus, the use of the proposed index to other cases requires the application of the Bayesian parameterization procedure to detect the optimal parameters. The parameterization of the index needs to be performed also when different satellite images are used (e.g. LANDSAT), due to their specific characteristics and spatial resolution”.

 After this article is published, future studies will take the proposed index and will apply it to other regions using Sentinel-2b or other satellite data with these bands.  These studies may not have access to high resolution image for verifying their output.  It is imperative for these authors (and anyone) who propose a new index to stress the assumptions/conditions under which it will work.

Action taken: The following sentence has been added in the Conclusion section (page 15, lines 518-524): “In this study, a reference image of actual reservoirs, obtained by manually digitalizing the reservoirs visible in high spatial resolution Google Earth images of the study area. Reservoirs in the area of study are affected by slight seasonal variations, therefore the reference map can be considered quite representative of the actual reservoirs in the catchment. It has to be remarked that, to ensure a reliable performance of the proposed procedure, the use of an accurate reference image for testing the effectiveness of the index is highly recommended when the reservoirs in the area of interest are subjected to seasonal or annual variations”.