the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 30 Mar 2021
Sea ice and water classification on dual-polarized Sentinel-1 imagery during melting season
Abstract. We provide a new sea ice and water classification product with high spatial and high temporal coverage using Sentinel-1 Synthetic Aperture Radar (SAR) data. The classification is applied in the Fram Strait region in the Arctic during melting seasons, when the contrast between backscatter intensities of different ice types observed by SAR is reduced due to the melted ice surface and wet snow on sea ice. The wet or melted snow strongly reduces the SAR penetration depth and thus suppresses the volume scattering contribution of sea ice. Furthermore, within the marginal sea ice zone (MIZ)
ambiguities between ice and water can result from the effects of winds and ocean currents on the ocean SAR backscatter.
On the other hand, under calm conditions the contrast between thin ice and flat open water can be reduced, and thus
decrease the separability of some ice. In summary, the melting season represents the most challenging time of the year for
reliable ice-water classification from SAR data. We propose here a new approach to overcome these problems by using a
mixture statistical distribution based conditional random fields (MSTA-CRF) model. To obtain reliable ice-water
classification whilst maintaining a fast computation time suitable for operational applications, the MSTA-CRF adopts a
superpixel approach in the fully connected CRF model. The MSTA-CRF is a semantic model, which integrates statistical
distributions (Gamma, Weibull, Alpha-Stable, etc.) to model the backscatters of ice and water and overcome the effects of
speckle noise and wind-roughened water. Dual-polarization Extended Wide (EW) mode Sentinel-1A/1B SAR data with
40 m spatial resolution is available several times per day within the Fram Strait region. Observations from June to
September during the six years 2015–2020 are collected and classified into ice and water categories. The classification
performance of algorithm is evaluated using ice charts from the Ice Service at the Norwegian Meteorological Institute
(MET Norway). The methods of training sample selection, and their application to processing large data volumes and
automatic classification of ice-water are discussed. In the experiment part, we demonstrate that the MSTA-CRF can provide
a good performance with about 90 % accuracy for ice-water classification, which is better than most of other state-of-the
art algorithms. Compared with the 89 GHz microwave radiometer ASI sea ice concentration product, the sea ice extent in
Fram Strait derived from MSTA-CRF algorithm is lower during melting seasons from 2015 to 2020, and the monthly June
to September sea ice area does not change so much in 2015–2017 and 2019–2020, but it has a significant decrease in 2018.
- Preprint
(6927 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on tc-2021-85', Anonymous Referee #1, 05 May 2021
Dear authors of the manuscript tc-2021-85,
In the manuscript a widely studied topic has been studied. The method is computationally quite
heavy but the results in classification are good. The manuscript is quite thorough, the
data set and the evaluation are quite comprehensive. There are still some aspects which need
to be taken into account before publishing this manuscript. In the following are my
comments.Test and training data set: It is not very clear how the data has been divided into
independent training and test data sets. Evaluation should be performed using a test data
set which is independent of the training data set, i.e. the training data set must be
excluded from the test data set. Now division into these two independent data sets is
not very clear to me. Please, in detail describe the division to independent training and test
(evaluation) data sets to confirm the reader that they are independent.Introduction:
Also sea ice concentration (SIC) estimates can be and are derived based on the proposed
SI/OW classification scheme. I recommend to include missing references to SAR-based
SIC estimation, there are many papers on this published during the recent years, e.g.:Wang, L., K. A. Scott, L. Xu, D. A. Clausi, Sea ice concentration estimation
during melt from dual-pol SAR scenes using deep convolutional
neural networks: A case study, IEEE Trans. Geosci. Remote Sens., vol.
54, no. 8, pp. 4524–4533, 2016.Wang, Scott, Clausi, Sea Ice Concentration Estimation during Freeze-Up from SAR Imagery
Using a Convolutional Neural Network Remote Sens. 2017, 9(5), 408; https://doi.org/10.3390/rs9050408W. Aldenhoff, A. Berg and L. E. B. Eriksson, "Sea ice concentration estimation from
Sentinel-1 Synthetic Aperture Radar images over the Fram Strait," 2016 IEEE International
Geoscience and Remote Sensing Symposium (IGARSS), 2016, pp. 7675-7677, doi: 10.1109/IGARSS.2016.7731001.Karvonen, Evaluation of the operational SAR based Baltic Sea ice concentration products,
Advances in Space Research 56(1), 2015, DOI: 10.1016/j.asr.2015.03.039And some references combining microwave radiometer and SAR for SIC estimation:
Karvonen, J., Baltic Sea Ice Concentration Estimation Using SENTINEL-1 SAR and AMSR2
Microwave Radiometer Data, IEEE Transactions on Geoscience and Remote Sensing (Volume: 55,
Issue: 5, May 2017), pp. 2871-2883, 2017, DOI: 10.1109/TGRS.2017.2655567.Malmgren-Hansen, D., Pedersen, L. T., Nielsen, A. A., Brandt Kreiner, M., Saldo, R.,
Skriver, H., Lavelle, J., Buus-Hinkler, J., Harnvig, K.,
A Convolutional Neural Network Architecture for Sentinel-1 and AMSR2 Data Fusion.
IEEE Transactions on Geoscience and Remote Sensing, v. 59, n. 3, pp. 1890-1902. 2021,
https://doi.org/10.1109/TGRS.2020.3004539Especially convolutional neural networks in sea ice classification and parameter estimation
have gained popularity during the recent years. These methods are computationally heavy
but software for their parallel efficient execution on graphics adapters exist.P3 L34: "...sea ice deformation features shows much more textual that decrease the
severability between flat thin ice and calm water." I don't understand this sentence. Please,
rewrite this.P4 2.1 Research area:
The sentence "To consider the spatial contextual information and preserve the spatial details
of each pixel in SAR imagery, the energy function based maximum a posteriori (MAP) estimation
in MSTA-CRF framework is proposed for operational ice water classification during melting
seasons in Fram Strait." does not belong to this subsection, it could be in introduction
or methodology section rather. Just start the section by "This study was performed in the
area of Fram Strait during the melting season." or something similar.P4 L21: "Figure 1 shows an overview of the research area and some satellite scenes used
in this manuscript." and Figure 1 / Figure 1 caption.
Why just some scenes are shown? Could the figure for example show the
total amount images at each location of the study area (by using some color coding), it would
be much more informative.P5 Sentinel-1 SAR Data:
L6: "... data during melting seasons from 2015 to 2020 are used.". Please, be more specific,
give the periods. Is the melting period the same every winter? E.g. some kind of temperature
statistics from nearby weather stations to confirm that the data represents melting period
every winter would be useful here.P6 Methodology:
Figure 2. If I have understood correctly SPAN image is used as an input? Now in the figure
there is an arrow from the leftmost SAR processing block to the MSTA-CRF block and it looks like
the uppermost row SAR data were input to the MSTA-CRF, possibly the arrow could be started from
the lower part of the block as is the second arrow. Assuming I have understood this correctly.P5 L5: SPAN of the HH and HV channels (sqrt(HH^2 + HV^2)). Later SPAN is defined as square root
of sigma0_HH^2 + sigma0_HV^2. Possibly the square root could be dropped from here and just say
that SPAN represent the joint total power of the two SAR channels and leave the more
precise definition later.P7 L16: There is wrong year in the publication, it should be 2017, not 2018.
There also exist this publication:
Park, Won, Korosov, Babiker, Miranda,
Textural Noise Correction for Sentinel-1 TOPSAR Cross-Polarization Channel Images,
IEEE Transactions on Geoscience and Remote Sensing ( Volume: 57, Issue: 6, June 2019),
DOI: 10.1109/TGRS.2018.2889381
Please, add the reference and check which version of the python SW have been used in this
study.P9 L8: "In corrected HV, backscatter coefficients of thin ice are greatly reduced."
This sentence is difficult (at least for me) to understand, Please, reformulate.P9 L13: "Figure 5 (e)", You probably mean Figure 4 (e)?
P9 L18 "for each category", as there are only two categories it would be better to say
"for both categories".P9 L19: "using the MET Norway ice charts". Please, be more specific and describe exactly how
the ice charts have been utilized.P10 L2: "each categories" -> "both categories"
P10 L10: "...that when the training samples reaches..."? Do you mean "...that when the
number of training samples reaches..."?P10 L32: "...contains a great number of scatters of radiation..." ->
"...contains a large number of scatterers of radiation...".P11 Table 3 caption: "...probability density function (pdf)..." -> "...probability density function (PDF)..."
or rather even "...probability density function..." and give the acronym PDF in the text.P11 L12: "PDF" -> "probability density function (PDF)", PDF always with capital letters.
P11 L22 and Eq. 8: Explain what is M (is it number of PDF's here?).
P12 L6: "...into several sub-superpixel using a random number..."? What does this mean?
"...into several sub-superpixel using a random number of pixels..."?P12 Eq. 9: What are K and N in the equation?
P12 "...y_i and y_j are the SAR backscatter coefficients at a pair of subsuperpixel..."? Are these
really SAR sigma0 values or SPAN values?P13 23: "CV (coefficient of variance)"? Do You mean "coefficient of variation"? At least for me
coefficient of variance is an unknown concept. If You use it, please, define it.P14 Fig. 6: Add x-axis labels ("model number" or something describing what is on the x-axis).
P14 Fig. 6: Be more specific in Y-axis label, now there is just "normalized". "normalized"
what? I guess "normalized parameter" would be better here. Fig. 6a is not very clear with so
many curves in one figure. Would there be any alternatives to make a more clear image (or
more than one image)?P15 L17-18: "PDF (probability density function)" This has already been opened on p. 11, so just write
"PDF".There seem to be some sentences which are not very easy to understand. I am not a native English
speaker and may not have noticed all of these sentences or possible grammar or typing errors.
I recommend to let a native Englsih speaker (Your co-author Nick Hughes) to check the sentences
and language of the revised manuscript before submission.Sincerely,
Citation: https://doi.org/10.5194/tc-2021-85-RC1 -
AC1: 'Reply on RC1', Yu Zhang, 28 Jun 2021
We are grateful to the reviewer for the constructive comments on our manuscript (tc-2021-85) entitled “Sea ice and water classification on dual-polarized Sentinel-1 imagery during melting season”. We have addressed all the comments. Our point-by-point responses are attached below in blue, while the original Reviewers’ comments are in black. You can find it in the supplement file.
Thank you again for valuable comments on our manuscript.
Sincerely,
-
AC1: 'Reply on RC1', Yu Zhang, 28 Jun 2021
-
RC2: 'Comment on tc-2021-85', Anonymous Referee #2, 13 May 2021
This manuscript introduces a new method for the ice mapping based on dual-polarized Sentinel-1 SAR data during summer season. The proposed method was developed from a conditional random field based on mixed statistical distribution. The results indicate the potential to derive reliable ice extent operationally. Unfortunately the author’s use of English is very poor, and the meaning was ambiguous and confusing in most cases. Missing methodological details, incorrect use of models, and the large number of typo and formatting errors do not make an impression of a self-contained manuscript. The authors should not submit a manuscript which is not ready for submission. I would recommend a rejection of this paper, but I think that the author could have a chance of publishing the results of their study if they prepare the manuscript better next time.
General Comments:
1. The description of the methodology is so poorly structured, which makes the logic of the research very confusing and hard to follow. One example is that the description of data preprocessing and training samples selection should not be introduced in the section of methodology. I am pretty sure a lot of efforts are still needed for improving the general structure of the paper.
2. The training samples was selected from the MET Norway ice chart. The MET ice chart is a weekly product and it inevitably has a time lag with SAR data. The change of sea ice is fast in melting period. How do you make sure the samples you choose are correct?
3. In the step of incidence angle correction, the authors used an incorrect sea ice scattering model. In equation (1), the backscattering of sea ice is described as the function of nadir backscattering and cosn(thetai). When the radar echo is incident vertically, the scattering mechanism of sea ice is specular scattering which is completely different from the scattering mechanism of SAR. Therefore, the used approach is illogical and unphysical.
4. The mean-shift method is critical to the proposed classification method. But the principle of mean-shift algorithm and the parameter setting for unsupervised segmentation should be introduced.
5. What I am most dissatisfied with is the use of distribution models. The distribution model of Gamma, Weibull and Alpha-stable is based on the statistical characteristics of pixels. However, the distribution model was for “sub-superpixel” (patches derived from unsupervised segmentation method) not for pixels. I don’t think these distribution models could be adaptable to image patches.
6. There are many SAR sea ice classification methods, taking these methods as baselines and comparing them with your method is necessary for validating the effect of your method. Moreover, the authors claimed that the advantage of proposed method is to identify sea ice in melting season. So you should give more examples to prove that the developed method can solve the problem of sea ice classification in summer.
7. According to the results of Table 4, the classification accuracy depends on used reference incidence angle. In equation (2), cosn(thetaref) is a constant value. I don’t understand why the variation of constant value has an impact on classification accuracy.
8. How to determine the parameters used in the proposed method (e.g. n and weight coefficients) is not clarified. Many details are not clear and need further explanation.
9. As the stated by the author, the accuracy of classification was validated by all the training data (see Page 10 Line 6). This is obviously incorrect. I am very confused about the sentence “If the overall accuracy (OA) is lower than 99%, we add 100 patches (50 for ice and 50 for water) from the rest of the training dataset to train the revised model, ……”. I’m not sure of your reasons for doing this?
Minor Comments:
Page 2, line 5: “search-and rescue” --> “search-and-rescue”.
Page 2, line 6: “ERS-1/-2, RADARSAT-1/-2, Sentinel-1A/-1B” --> “ERS-1/2, RADARSAT-1/2, Sentinel-1A/B”.
Page 2, line 14: “introduced” --> “have”.
Page 2, line 15: “channel” --> “polarization”. Please replace “channel” with “polarization” in the full text.
Page 2, line 18: “for improved” --> “for improving”.
Page 2, line 27: here Radarsat-2 is “RS-2”, but its abbreviation is “RS2” in line 6.
Page 2, line 28: “-3” --> “Sentinel-3”.
Page 2, line 29: “with low resolution passive microwave form low resolution microwave from AMSR2” reformulate this sentence.
Page 2, line 32: “As the backscattering is usually affected by ocean waves propagating into the ice area, ….” for thin sea ice, the backscattering coefficient could be affected by wave. But for thick sea ice, the effect of waves on backscattering is very low.
Page 3, line 7: “SVM realize” --> “realizes”.
Page 3, line 7: “by training the kernel with the transformation into high dimensional space,” reformulate this sentence.
Page 3, line 9: “Textual feature based neural network methods also shows” --> “show”.
Page 3, line 10: “Murashkin et al. (2018) use” --> “used”.
Page 3, line 11: “th MIZ” --> “the MIZ”.
I stop here with my comments and I think I almost had comments in every single sentence. There are a lot of grammatical issues but also, more seriously, inaccurate statements.
Citation: https://doi.org/10.5194/tc-2021-85-RC2 -
AC5: 'Reply on RC2', Yu Zhang, 01 Jul 2021
We are grateful to the reviewer for the constructive comments on our manuscript (tc-2021-85) entitled “Sea ice and water classification on dual-polarized Sentinel-1 imagery during melting season”. We have addressed all the comments. Our point-by-point responses are attached below in blue, while the original Reviewers’ comments are in black. You can find it in the supplement file.
Thank you again for valuable comments on our manuscript.
Sincerely,
-
AC5: 'Reply on RC2', Yu Zhang, 01 Jul 2021
-
RC3: 'Comment on tc-2021-85', Anonymous Referee #3, 22 May 2021
Review of tc-2021-85
Submission to The Cryosphere
Summary
The authors detail a SAR image sea ice-water classification technique for use during melting conditions in the Fram Strait region. The input data are dual-polarization (HH + HV) Sentinel-1 EW mode scenes which are widely available over marine regions and open access. Their method for pre-processing of the HV channel of the Sentinel-1 SAR data seems to work very well, enabling its inclusion in the classifier. A good classification accuracy of ~90% is achieved, and the results are used to examine sea ice concentration evolution in the summer months over the 2015-2020 period. Since C-band SAR images are commonly used for ice mapping and charting, the results are potentially extendable to other missions as well. The potential to use a SAR based sea ice concentration algorithm during the summer months, and in a marginal ice zone, when/where passive microwave data is less reliable, is also noteworthy.
- The paper is hard to follow, especially given that there is a lot of repetition in the text and figures, and some concepts and acronyms defined more than once. The input data to classification result is shown in Fig. 2 and Fig. 5; the SAR processing to remove noise is shown in Figs. 2, 3, and 4. Training sample selection is detailed in Sections 3.2. and 4.2. CRF and MSTA-CRF are defined on Page 3 then defined again on Page 6 (etc.). The selection of reference incidence of 23° doesn’t need to be introduced on Page 8 then again on Page 13. The authors should describe their methodology in terms of input data, pre-processing, training, classification, and validation, and make it shorter in length. Everything on Page 17 and later could be included in Results and Discussion.
- It is unclear what input data is actually used. Fig. 1 shows some scene extents though it is difficult to tell whether they are arbitrarily chosen or what they are supposed to represent. Later in the paper there is mention of 488 images, or one image each day from June to September over the period of 2015-2018. Provide more detail on what Sentinel-1 data are used (without listing them).
- The images are described as pertaining to melting conditions. More justification for this should be provided since it is insufficient to assume that all images between June and September are in melting conditions at this latitude.
- If there is a Sentinel-1 image from each day in the Fram Strait, images that correspond more closely to the MET Norway ice charts should be used for selection of training data. Otherwise there is more chance for ice drift and changing ice/water conditions to introduce error into the training sample selection.
- Does the inclusion of GLCM features in the SVM classifier improve its performance when compared to using just HH + HV data? The inclusion of GLCM features is described but it is unclear why, and on what basis the GLCM parameters, the kernel size, quantization level, and displacement were chosen
- The main misclassification error, on a class-by-class basis, is given to be caused by the presence of melting water on fast ice, leading to misclassification of ice as open water. However it is unclear how this was determined. If it is assumed, then the authors should provide some justification for it (e.g. article reference).
- Consistency in terminology is needed, e.g. “backscatter”, “backscatters”, “backscattering”, and “backscatter coefficient”; “incidence angle” and “incident angle”; “RS-2” and “RS2” etc.
Specific comments:
(Page = P, Line = L)
P1L23: “backscatters” should be “backscatter”
P2L21: delete “value”
P2L23: should be “MAp-Guided”
P2L30: data “are” (plural)
P2L32: Backscatter is also affected by waves that form, e.g. by capillary action, not just waves propagating into the area.
P3L2: Delete “scatters of the”; also change “the mixture” to “a mixture” on the next line.
P3L7: “analysis” should be “classification”
P3L9: “Texture”
P3L14: Use of the term “usually” here creates ambiguity.
P3L26: It would be helpful to more clearly define superpixel and sub-superpixel at their first use.
P3L31: “A statistical distribution ….”
P3L34: Some spelling mistakes here.
P4L15-18: There is a lot of detail given about the classification method here. The focus should be on Fram Strait.
Fig. 1.: Make a better map with the image detail provided.
P5L6: “mode”
P5L15: Very Open Drift is defined as SIC<1 here, whereas in Table 2 it is shown as 1-4.
P7L12: “sub-swaths”
P7L18-19: Delete sentence that starts “Preprocessing methods …”
P8L9-10: Delete the description “with SPAN being defined…” etc. since the equation is given. The equation doesn’t need to be in Fig. 4.
P10L3: Delete “e.g.” and correct “otherwise”
P10L32: “noise is based on…”
P12L8: Sentence beginning “It may have” is hard to understand. Perhaps break it up.
P13L28: The table isn’t really necessary since the analysis and its outcome is described well above it.
P15L31: “and Weibull distributions are not …”
P17L10: Delete “In the experiment”
P18L12: Provide some detail about the temporal offset between the classification result and the ice chart.
P20L13: “MSTA”
P21L4: “from the same orbits”
P22L3: delete “has”
Citation: https://doi.org/10.5194/tc-2021-85-RC3 -
AC3: 'Reply on RC3', Yu Zhang, 28 Jun 2021
We are grateful to the reviewer for the constructive comments on our manuscript (tc-2021-85) entitled “Sea ice and water classification on dual-polarized Sentinel-1 imagery during melting season”. We have addressed all the comments. Our point-by-point responses are attached below in blue, while the original Reviewers’ comments are in black. You can find it in the supplement file.
Thank you again for valuable comments on our manuscript.
Sincerely,
Status: closed
-
RC1: 'Comment on tc-2021-85', Anonymous Referee #1, 05 May 2021
Dear authors of the manuscript tc-2021-85,
In the manuscript a widely studied topic has been studied. The method is computationally quite
heavy but the results in classification are good. The manuscript is quite thorough, the
data set and the evaluation are quite comprehensive. There are still some aspects which need
to be taken into account before publishing this manuscript. In the following are my
comments.Test and training data set: It is not very clear how the data has been divided into
independent training and test data sets. Evaluation should be performed using a test data
set which is independent of the training data set, i.e. the training data set must be
excluded from the test data set. Now division into these two independent data sets is
not very clear to me. Please, in detail describe the division to independent training and test
(evaluation) data sets to confirm the reader that they are independent.Introduction:
Also sea ice concentration (SIC) estimates can be and are derived based on the proposed
SI/OW classification scheme. I recommend to include missing references to SAR-based
SIC estimation, there are many papers on this published during the recent years, e.g.:Wang, L., K. A. Scott, L. Xu, D. A. Clausi, Sea ice concentration estimation
during melt from dual-pol SAR scenes using deep convolutional
neural networks: A case study, IEEE Trans. Geosci. Remote Sens., vol.
54, no. 8, pp. 4524–4533, 2016.Wang, Scott, Clausi, Sea Ice Concentration Estimation during Freeze-Up from SAR Imagery
Using a Convolutional Neural Network Remote Sens. 2017, 9(5), 408; https://doi.org/10.3390/rs9050408W. Aldenhoff, A. Berg and L. E. B. Eriksson, "Sea ice concentration estimation from
Sentinel-1 Synthetic Aperture Radar images over the Fram Strait," 2016 IEEE International
Geoscience and Remote Sensing Symposium (IGARSS), 2016, pp. 7675-7677, doi: 10.1109/IGARSS.2016.7731001.Karvonen, Evaluation of the operational SAR based Baltic Sea ice concentration products,
Advances in Space Research 56(1), 2015, DOI: 10.1016/j.asr.2015.03.039And some references combining microwave radiometer and SAR for SIC estimation:
Karvonen, J., Baltic Sea Ice Concentration Estimation Using SENTINEL-1 SAR and AMSR2
Microwave Radiometer Data, IEEE Transactions on Geoscience and Remote Sensing (Volume: 55,
Issue: 5, May 2017), pp. 2871-2883, 2017, DOI: 10.1109/TGRS.2017.2655567.Malmgren-Hansen, D., Pedersen, L. T., Nielsen, A. A., Brandt Kreiner, M., Saldo, R.,
Skriver, H., Lavelle, J., Buus-Hinkler, J., Harnvig, K.,
A Convolutional Neural Network Architecture for Sentinel-1 and AMSR2 Data Fusion.
IEEE Transactions on Geoscience and Remote Sensing, v. 59, n. 3, pp. 1890-1902. 2021,
https://doi.org/10.1109/TGRS.2020.3004539Especially convolutional neural networks in sea ice classification and parameter estimation
have gained popularity during the recent years. These methods are computationally heavy
but software for their parallel efficient execution on graphics adapters exist.P3 L34: "...sea ice deformation features shows much more textual that decrease the
severability between flat thin ice and calm water." I don't understand this sentence. Please,
rewrite this.P4 2.1 Research area:
The sentence "To consider the spatial contextual information and preserve the spatial details
of each pixel in SAR imagery, the energy function based maximum a posteriori (MAP) estimation
in MSTA-CRF framework is proposed for operational ice water classification during melting
seasons in Fram Strait." does not belong to this subsection, it could be in introduction
or methodology section rather. Just start the section by "This study was performed in the
area of Fram Strait during the melting season." or something similar.P4 L21: "Figure 1 shows an overview of the research area and some satellite scenes used
in this manuscript." and Figure 1 / Figure 1 caption.
Why just some scenes are shown? Could the figure for example show the
total amount images at each location of the study area (by using some color coding), it would
be much more informative.P5 Sentinel-1 SAR Data:
L6: "... data during melting seasons from 2015 to 2020 are used.". Please, be more specific,
give the periods. Is the melting period the same every winter? E.g. some kind of temperature
statistics from nearby weather stations to confirm that the data represents melting period
every winter would be useful here.P6 Methodology:
Figure 2. If I have understood correctly SPAN image is used as an input? Now in the figure
there is an arrow from the leftmost SAR processing block to the MSTA-CRF block and it looks like
the uppermost row SAR data were input to the MSTA-CRF, possibly the arrow could be started from
the lower part of the block as is the second arrow. Assuming I have understood this correctly.P5 L5: SPAN of the HH and HV channels (sqrt(HH^2 + HV^2)). Later SPAN is defined as square root
of sigma0_HH^2 + sigma0_HV^2. Possibly the square root could be dropped from here and just say
that SPAN represent the joint total power of the two SAR channels and leave the more
precise definition later.P7 L16: There is wrong year in the publication, it should be 2017, not 2018.
There also exist this publication:
Park, Won, Korosov, Babiker, Miranda,
Textural Noise Correction for Sentinel-1 TOPSAR Cross-Polarization Channel Images,
IEEE Transactions on Geoscience and Remote Sensing ( Volume: 57, Issue: 6, June 2019),
DOI: 10.1109/TGRS.2018.2889381
Please, add the reference and check which version of the python SW have been used in this
study.P9 L8: "In corrected HV, backscatter coefficients of thin ice are greatly reduced."
This sentence is difficult (at least for me) to understand, Please, reformulate.P9 L13: "Figure 5 (e)", You probably mean Figure 4 (e)?
P9 L18 "for each category", as there are only two categories it would be better to say
"for both categories".P9 L19: "using the MET Norway ice charts". Please, be more specific and describe exactly how
the ice charts have been utilized.P10 L2: "each categories" -> "both categories"
P10 L10: "...that when the training samples reaches..."? Do you mean "...that when the
number of training samples reaches..."?P10 L32: "...contains a great number of scatters of radiation..." ->
"...contains a large number of scatterers of radiation...".P11 Table 3 caption: "...probability density function (pdf)..." -> "...probability density function (PDF)..."
or rather even "...probability density function..." and give the acronym PDF in the text.P11 L12: "PDF" -> "probability density function (PDF)", PDF always with capital letters.
P11 L22 and Eq. 8: Explain what is M (is it number of PDF's here?).
P12 L6: "...into several sub-superpixel using a random number..."? What does this mean?
"...into several sub-superpixel using a random number of pixels..."?P12 Eq. 9: What are K and N in the equation?
P12 "...y_i and y_j are the SAR backscatter coefficients at a pair of subsuperpixel..."? Are these
really SAR sigma0 values or SPAN values?P13 23: "CV (coefficient of variance)"? Do You mean "coefficient of variation"? At least for me
coefficient of variance is an unknown concept. If You use it, please, define it.P14 Fig. 6: Add x-axis labels ("model number" or something describing what is on the x-axis).
P14 Fig. 6: Be more specific in Y-axis label, now there is just "normalized". "normalized"
what? I guess "normalized parameter" would be better here. Fig. 6a is not very clear with so
many curves in one figure. Would there be any alternatives to make a more clear image (or
more than one image)?P15 L17-18: "PDF (probability density function)" This has already been opened on p. 11, so just write
"PDF".There seem to be some sentences which are not very easy to understand. I am not a native English
speaker and may not have noticed all of these sentences or possible grammar or typing errors.
I recommend to let a native Englsih speaker (Your co-author Nick Hughes) to check the sentences
and language of the revised manuscript before submission.Sincerely,
Citation: https://doi.org/10.5194/tc-2021-85-RC1 -
AC1: 'Reply on RC1', Yu Zhang, 28 Jun 2021
We are grateful to the reviewer for the constructive comments on our manuscript (tc-2021-85) entitled “Sea ice and water classification on dual-polarized Sentinel-1 imagery during melting season”. We have addressed all the comments. Our point-by-point responses are attached below in blue, while the original Reviewers’ comments are in black. You can find it in the supplement file.
Thank you again for valuable comments on our manuscript.
Sincerely,
-
AC1: 'Reply on RC1', Yu Zhang, 28 Jun 2021
-
RC2: 'Comment on tc-2021-85', Anonymous Referee #2, 13 May 2021
This manuscript introduces a new method for the ice mapping based on dual-polarized Sentinel-1 SAR data during summer season. The proposed method was developed from a conditional random field based on mixed statistical distribution. The results indicate the potential to derive reliable ice extent operationally. Unfortunately the author’s use of English is very poor, and the meaning was ambiguous and confusing in most cases. Missing methodological details, incorrect use of models, and the large number of typo and formatting errors do not make an impression of a self-contained manuscript. The authors should not submit a manuscript which is not ready for submission. I would recommend a rejection of this paper, but I think that the author could have a chance of publishing the results of their study if they prepare the manuscript better next time.
General Comments:
1. The description of the methodology is so poorly structured, which makes the logic of the research very confusing and hard to follow. One example is that the description of data preprocessing and training samples selection should not be introduced in the section of methodology. I am pretty sure a lot of efforts are still needed for improving the general structure of the paper.
2. The training samples was selected from the MET Norway ice chart. The MET ice chart is a weekly product and it inevitably has a time lag with SAR data. The change of sea ice is fast in melting period. How do you make sure the samples you choose are correct?
3. In the step of incidence angle correction, the authors used an incorrect sea ice scattering model. In equation (1), the backscattering of sea ice is described as the function of nadir backscattering and cosn(thetai). When the radar echo is incident vertically, the scattering mechanism of sea ice is specular scattering which is completely different from the scattering mechanism of SAR. Therefore, the used approach is illogical and unphysical.
4. The mean-shift method is critical to the proposed classification method. But the principle of mean-shift algorithm and the parameter setting for unsupervised segmentation should be introduced.
5. What I am most dissatisfied with is the use of distribution models. The distribution model of Gamma, Weibull and Alpha-stable is based on the statistical characteristics of pixels. However, the distribution model was for “sub-superpixel” (patches derived from unsupervised segmentation method) not for pixels. I don’t think these distribution models could be adaptable to image patches.
6. There are many SAR sea ice classification methods, taking these methods as baselines and comparing them with your method is necessary for validating the effect of your method. Moreover, the authors claimed that the advantage of proposed method is to identify sea ice in melting season. So you should give more examples to prove that the developed method can solve the problem of sea ice classification in summer.
7. According to the results of Table 4, the classification accuracy depends on used reference incidence angle. In equation (2), cosn(thetaref) is a constant value. I don’t understand why the variation of constant value has an impact on classification accuracy.
8. How to determine the parameters used in the proposed method (e.g. n and weight coefficients) is not clarified. Many details are not clear and need further explanation.
9. As the stated by the author, the accuracy of classification was validated by all the training data (see Page 10 Line 6). This is obviously incorrect. I am very confused about the sentence “If the overall accuracy (OA) is lower than 99%, we add 100 patches (50 for ice and 50 for water) from the rest of the training dataset to train the revised model, ……”. I’m not sure of your reasons for doing this?
Minor Comments:
Page 2, line 5: “search-and rescue” --> “search-and-rescue”.
Page 2, line 6: “ERS-1/-2, RADARSAT-1/-2, Sentinel-1A/-1B” --> “ERS-1/2, RADARSAT-1/2, Sentinel-1A/B”.
Page 2, line 14: “introduced” --> “have”.
Page 2, line 15: “channel” --> “polarization”. Please replace “channel” with “polarization” in the full text.
Page 2, line 18: “for improved” --> “for improving”.
Page 2, line 27: here Radarsat-2 is “RS-2”, but its abbreviation is “RS2” in line 6.
Page 2, line 28: “-3” --> “Sentinel-3”.
Page 2, line 29: “with low resolution passive microwave form low resolution microwave from AMSR2” reformulate this sentence.
Page 2, line 32: “As the backscattering is usually affected by ocean waves propagating into the ice area, ….” for thin sea ice, the backscattering coefficient could be affected by wave. But for thick sea ice, the effect of waves on backscattering is very low.
Page 3, line 7: “SVM realize” --> “realizes”.
Page 3, line 7: “by training the kernel with the transformation into high dimensional space,” reformulate this sentence.
Page 3, line 9: “Textual feature based neural network methods also shows” --> “show”.
Page 3, line 10: “Murashkin et al. (2018) use” --> “used”.
Page 3, line 11: “th MIZ” --> “the MIZ”.
I stop here with my comments and I think I almost had comments in every single sentence. There are a lot of grammatical issues but also, more seriously, inaccurate statements.
Citation: https://doi.org/10.5194/tc-2021-85-RC2 -
AC5: 'Reply on RC2', Yu Zhang, 01 Jul 2021
We are grateful to the reviewer for the constructive comments on our manuscript (tc-2021-85) entitled “Sea ice and water classification on dual-polarized Sentinel-1 imagery during melting season”. We have addressed all the comments. Our point-by-point responses are attached below in blue, while the original Reviewers’ comments are in black. You can find it in the supplement file.
Thank you again for valuable comments on our manuscript.
Sincerely,
-
AC5: 'Reply on RC2', Yu Zhang, 01 Jul 2021
-
RC3: 'Comment on tc-2021-85', Anonymous Referee #3, 22 May 2021
Review of tc-2021-85
Submission to The Cryosphere
Summary
The authors detail a SAR image sea ice-water classification technique for use during melting conditions in the Fram Strait region. The input data are dual-polarization (HH + HV) Sentinel-1 EW mode scenes which are widely available over marine regions and open access. Their method for pre-processing of the HV channel of the Sentinel-1 SAR data seems to work very well, enabling its inclusion in the classifier. A good classification accuracy of ~90% is achieved, and the results are used to examine sea ice concentration evolution in the summer months over the 2015-2020 period. Since C-band SAR images are commonly used for ice mapping and charting, the results are potentially extendable to other missions as well. The potential to use a SAR based sea ice concentration algorithm during the summer months, and in a marginal ice zone, when/where passive microwave data is less reliable, is also noteworthy.
- The paper is hard to follow, especially given that there is a lot of repetition in the text and figures, and some concepts and acronyms defined more than once. The input data to classification result is shown in Fig. 2 and Fig. 5; the SAR processing to remove noise is shown in Figs. 2, 3, and 4. Training sample selection is detailed in Sections 3.2. and 4.2. CRF and MSTA-CRF are defined on Page 3 then defined again on Page 6 (etc.). The selection of reference incidence of 23° doesn’t need to be introduced on Page 8 then again on Page 13. The authors should describe their methodology in terms of input data, pre-processing, training, classification, and validation, and make it shorter in length. Everything on Page 17 and later could be included in Results and Discussion.
- It is unclear what input data is actually used. Fig. 1 shows some scene extents though it is difficult to tell whether they are arbitrarily chosen or what they are supposed to represent. Later in the paper there is mention of 488 images, or one image each day from June to September over the period of 2015-2018. Provide more detail on what Sentinel-1 data are used (without listing them).
- The images are described as pertaining to melting conditions. More justification for this should be provided since it is insufficient to assume that all images between June and September are in melting conditions at this latitude.
- If there is a Sentinel-1 image from each day in the Fram Strait, images that correspond more closely to the MET Norway ice charts should be used for selection of training data. Otherwise there is more chance for ice drift and changing ice/water conditions to introduce error into the training sample selection.
- Does the inclusion of GLCM features in the SVM classifier improve its performance when compared to using just HH + HV data? The inclusion of GLCM features is described but it is unclear why, and on what basis the GLCM parameters, the kernel size, quantization level, and displacement were chosen
- The main misclassification error, on a class-by-class basis, is given to be caused by the presence of melting water on fast ice, leading to misclassification of ice as open water. However it is unclear how this was determined. If it is assumed, then the authors should provide some justification for it (e.g. article reference).
- Consistency in terminology is needed, e.g. “backscatter”, “backscatters”, “backscattering”, and “backscatter coefficient”; “incidence angle” and “incident angle”; “RS-2” and “RS2” etc.
Specific comments:
(Page = P, Line = L)
P1L23: “backscatters” should be “backscatter”
P2L21: delete “value”
P2L23: should be “MAp-Guided”
P2L30: data “are” (plural)
P2L32: Backscatter is also affected by waves that form, e.g. by capillary action, not just waves propagating into the area.
P3L2: Delete “scatters of the”; also change “the mixture” to “a mixture” on the next line.
P3L7: “analysis” should be “classification”
P3L9: “Texture”
P3L14: Use of the term “usually” here creates ambiguity.
P3L26: It would be helpful to more clearly define superpixel and sub-superpixel at their first use.
P3L31: “A statistical distribution ….”
P3L34: Some spelling mistakes here.
P4L15-18: There is a lot of detail given about the classification method here. The focus should be on Fram Strait.
Fig. 1.: Make a better map with the image detail provided.
P5L6: “mode”
P5L15: Very Open Drift is defined as SIC<1 here, whereas in Table 2 it is shown as 1-4.
P7L12: “sub-swaths”
P7L18-19: Delete sentence that starts “Preprocessing methods …”
P8L9-10: Delete the description “with SPAN being defined…” etc. since the equation is given. The equation doesn’t need to be in Fig. 4.
P10L3: Delete “e.g.” and correct “otherwise”
P10L32: “noise is based on…”
P12L8: Sentence beginning “It may have” is hard to understand. Perhaps break it up.
P13L28: The table isn’t really necessary since the analysis and its outcome is described well above it.
P15L31: “and Weibull distributions are not …”
P17L10: Delete “In the experiment”
P18L12: Provide some detail about the temporal offset between the classification result and the ice chart.
P20L13: “MSTA”
P21L4: “from the same orbits”
P22L3: delete “has”
Citation: https://doi.org/10.5194/tc-2021-85-RC3 -
AC3: 'Reply on RC3', Yu Zhang, 28 Jun 2021
We are grateful to the reviewer for the constructive comments on our manuscript (tc-2021-85) entitled “Sea ice and water classification on dual-polarized Sentinel-1 imagery during melting season”. We have addressed all the comments. Our point-by-point responses are attached below in blue, while the original Reviewers’ comments are in black. You can find it in the supplement file.
Thank you again for valuable comments on our manuscript.
Sincerely,
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,357 | 915 | 68 | 2,340 | 60 | 58 |
- HTML: 1,357
- PDF: 915
- XML: 68
- Total: 2,340
- BibTeX: 60
- EndNote: 58
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Cited
5 citations as recorded by crossref.
- Deep Learning Based Sea Ice Classification with Gaofen-3 Fully Polarimetric SAR Data T. Zhang et al. 10.3390/rs13081452
- River ice monitoring of the Danube and Tisza rivers using Sentinel-1 radar data L. van et al. 10.5937/gp26-39962
- MMSeaIce: a collection of techniques for improving sea ice mapping with a multi-task model X. Chen et al. 10.5194/tc-18-1621-2024
- Sea Ice–Water Classification of RADARSAT-2 Imagery Based on Residual Neural Networks (ResNet) with Regional Pooling M. Jiang et al. 10.3390/rs14133025
- Uncertainty-Incorporated Ice and Open Water Detection on Dual-Polarized SAR Sea Ice Imagery X. Chen et al. 10.1109/TGRS.2022.3233871