Classification of Nemoral Forests with Fusion of Multi-Temporal Sentinel-1 and 2 Data

Round 1

Reviewer 1 Report

Dear Editors:

In the manuscript the forest land, forest type and tree species groups of nemoreal forests are classified applying NFI plots and multitemporal Sentinel 1 and 2 satellite data. The study covers the whole Denmark.
The classification method is well known random forest.

My general comments are:

The topic of the manuscript is well aligned with the data needs of greenhouse gas inventory and forest resrources assessment. Obviously the GEE enables using a large set of various remote sensing data covering large areas.

However, the manuscript lacks some essential presentation, e.g. about the Sentinel image material (calibration, rectification errors), and in general how the multitemporal and multisensor image mosaic actually is formed and what data is used for a single pixel.
The Random forest method is very briefly described only. The usual image analysis feature selection and standardization steps are not dealed.
Overall, a standard way classification is used to a large RS data set and a sampling based large field plot data set in an efficient way, but apart from
that there is not so much novelty in the methods.

Detailed comments:

Lines 66-79: Introduction is on a general level, no details of the used
classification methods mentioned for the articles refered.

Line 107: BOA atmospheric correction used? The accuracy ot the method?

Line 108-109: Automatic cloud-masking tend to be unreliable; was there any visual checking applied?

Lines 122-123: Time series or Image mosaics? The size of SAR image frames?
Same question for Sentinel-2 images.
It should be made very clear, how many temporal sets there are for
each sensor and how many image frames to needed to cover the test area.
This all remains unclear now.

Lines 178-179: What about in the validation/test data? If the ambiguous plots
are not included may give optimistic accuracies of classification, because a
full cover classification will have to be carried out to make a land cover map.

Lines 180-182: Why not use a forest type 'mixed' also? If plenty of plots from the target set are omitted and classified only with a subset of plots the true value of accuracies obtained may be questionable.

Line 212-213: Were the changed plots between field measurement date and
image acquisition date checked and removed if necessary?

Line 233-234: What about the image features, any standardisation used for them? Or subselection of feature set?

Lines 238-239: Please explain briefly this "good practice" so that the readers don't have to find themselves the methods used.

Lines 289-291: The producer's accuracies are for the selected subset. But to obtain the true accuracy (mapping the forests), should the all the plots, also ambiguous ones, be inluded?

Lines 292: How did you adjust the error to the estimate?

Lines 348-349: Did you trye the classification with a monotemporal image set?

Lines 380-381: The OA is sensitive to the proportion of forest/non-forest areas; with a one dominating class it is easy to obtain high OAs.

Line 398-401: Are the accuracies comparable? Here the mixed forest plots were censored from the data which will probably give rise in the accuracy.

Lines 404-405: How do you define data fusion in remote sensing research context?

Lines 437-438: Would this be for S-2 optical images? But the best feature in your study was SAR image feature.

Line 453: The rectified pixel size was 10 m but what were the actual pixel sizes in the source images?

Author Response

We want to thank both reviewers for their constructive comments aiding us in our work to improve the manuscript and making it suitable for publication in Remote Sensing.

Our detailed response to each of the reviewers' suggestions may be found in the attached pdf-document.

Please note that our response to both reviewer’s comments are submitted to each reviewer, since some comments made by the two reviewers are interrelated. Hereby, we hope to provide both reviewers with a full overview of the changes made to the manuscript. A detailed record of our response to each of the suggestions made by each reviewer is found in different section labelled “Reviewer #1” and “Reviewer #2”.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

This is a suitable manuscript for this journal

Please choose a better title for this manuscript

Instruction section must be developed. Current form is not acceptable, please describe more about propose of study and also have better literature review.

This manuscript had two Figure1 ! Please correct them and follow journal’s format.

Table 2 is required more explanation.

In result section and before Discussion section have a final figure about result, like a figure by total view. Conclusions must be developed. Please add some new conclusion, no need repeat again the results.

Use some new relevant references.

Author Response

We want to thank both reviewers for their constructive comments aiding us in our work to improve the manuscript and making it suitable for publication in Remote Sensing.

Our detailed response to the suggestions made by the reviwers' comments may be found in the attached pdf-document.

Please note that our response to both reviewer’s comments are submitted to each reviewer, since some comments made by the two reviewers are interrelated. Hereby, we hope to provide both reviewers with a full overview of the changes made to the manuscript. A detailed record of our response to each of the suggestions made by each reviewer is found in different section labelled “Reviewer #1” and “Reviewer #2”.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have improved the manuscript substantially. The introduction is well written now. Also the explanation on the image processing is much clearer now.

Lines 298-305:

Even though selected grid cell size is 10x10m2 the original sensor resolution 
(Sentinel-2) is larger for some channels and in addition the sensor point distribution function typically is larger than the nominal pixel size thus gathering reflectance from even larger areas. 
Also, if  median functions are applied to several images over same pixel, would it not be affected by the locational errors in between images and this  would add spatial variation to the output image.
For the NFI plots, of course it is relevant to select good observations for the training data. 
But one could use all the plots for target data although the large size of the 
SSU plot sets limitations, e.g. plots partly on other land, and some compromises are needed.
Compare this to the situation when a classified map should be 
validated by selecting a statistical sample of plots. In this case all 
the points or pixels within the forest map (the population) should have 
a probability to be selected to the sample.

Author Response

Dear reviewer,

Firstly, let me take this opportunity to thank you for your meticulous effort to improve our manuscript and make it suitable for publication in Remote Sensing. We have now carefully revised the manuscript in accordance with your instructions. Please find below, the specific changes made to the manuscript.

Please note that we attached a manuscript with track changes where you may assess the specific changes made to the manuscript. Please also note that our response to both reviewer’s comments are submitted to each reviewer, since some comments made by the two reviewers are interrelated. Hereby, we hope to provide both reviewers with a full overview of the changes made to the manuscript.

Yours,

Thomas Nord-Larsen

 

Reviewer #1

The authors have improved the manuscript substantially. The introduction is well written now. Also the explanation on the image processing is much clearer now.

Lines 298-305:

Even though selected grid cell size is 10x10m2 the original sensor resolution (Sentinel-2) is larger for some channels and in addition the sensor point distribution function typically is larger than the nominal pixel size thus gathering reflectance from even larger areas. Also, if median functions are applied to several images over same pixel, would it not be affected by the locational errors in between images and this  would add spatial variation to the output image.

Reviewer is right that accuracy of S-2 geo-referencing should be mentioned and kept in mind when doing such type of analysis. In the revised MS we added (l: 152-155):

After the activation of the Global Reference Image (GRI) geometric refinement, the geometric accuracy of S-2 multi-temporal co-registration is now better than 0.3 pixel (at 2σ confidence level) [48].

48. Gascon, Ferran, Catherine Bouzinac, Olivier Thépaut, Mathieu Jung, Benjamin Francesconi, Jérôme Louis, Vincent Lonjou, Bruno Lafrance, Stéphane Massera, Angélique Gaudel-Vacaresse, Florie Languille, Bahjat Alhammoud, Françoise Viallefont, Bringfried Pflug, Jakub Bieniarz, Sébastien Clerc, Laëtitia Pessiot, Thierry Trémas, Enrico Cadau, Roberto De Bonis, Claudia Isola, Philippe Martimort, and Valérie Fernandez. "Copernicus Sentinel-2a Calibration and Products Validation Status." Remote Sensing 9, no. 6 (2017): 584.

For the NFI plots, of course it is relevant to select good observations for the training data. But one could use all the plots for target data although the large size of the SSU plot sets limitations, e.g. plots partly on other land, and some compromises are needed. Compare this to the situation when a classified map should be validated by selecting a statistical sample of plots. In this case all the points or pixels within the forest map (the population) should have a probability to be selected to the sample.

In response to the reviwers comment, we included a description of the more common procedure of obtaining ground truth from e.g. photointerpretation with a finer resolution than the resulting map, which allows unique classification of each individual pixel. We further included references to extensive reviews, dealing with these matters (l: 239-247):

“The use of NFI sample plots for pixel-based classification of remotely sensed imagery may be problematic when labelling is obtained from the 706 m2 large circular sample plots, but the classification is subsequently trained and evaluated on 100 m2 (10x10 m) pixels [55]. In this case, the labelling obtained from the entire plot may not apply to the individual pixels encompassed (i.e. when one sample plot spans several land cover classes), leading to poorer training of the classifier. This is contrasting the common practice of selecting a statistical sample for classification or evaluation from an already classified map commonly derived from data with similar or finer resolution than the resulting map, where all pixels may be uniquely classified [56].”

55. Foody, Giles M. "Status of Land Cover Classification Accuracy Assessment." Remote Sensing of Environment 80, no. 1 (2002): 185-201.
56. Stehman, Stephen V., and Giles M. Foody. "Key Issues in Rigorous Accuracy Assessment of Land Cover Products." Remote Sensing of Environment 231 (2019): 111199.

 

Reviewer #2

The author improved manuscript accordingly. Conclusion Section still needs a major revision.

Please re-write it. No need repeat again the result in conclusion. Please mention your finding and provide some recommendation and limitation for this study.

In the revised manuscript we already provided recommendations for future mapping efforts, i.e. that the use of multi-temporal images yields robust estimates and that these may be improved by a finer temporal scale. However, in response to the comments also posed by reviewer #1, we included a reference to the limitation related to the use of NFI data for training and validation:

“Albeit sample plots with ambiguous land-use or tree species characteristics posed a specific challenge in relation to selecting samples for training and validation, the NFI data provided an unbiased sample covering the entire domain, controlled in the field, and readily available to the user.”

According to normal scientific practice, it should be possible to read the conclusion separate from the main document. Consequently, our results must be presented briefly and related to the scientific development, as was already done in the revised manuscript. We are thus reluctant to rewrite the conclusions, leaving out the results of our study as otherwise suggested by the reviewer, as we believe that this would impair the readability of the conclusion. We encourage the editor to provide directions as to whether we should leave out results entirely from the Conclusion-section.

Author Response File: Author Response.docx

Reviewer 2 Report

The author improved manuscript accordingly. Conclusion Section still needs a major revision.

Please re-write it. No need repeat again the result in conclusion. Please mention your finding and provide some recommendation and limitation for this study.

Author Response

Dear reviewer,

Firstly, let me take this opportunity to thank you for your meticulous effort to improve our manuscript and make it suitable for publication in Remote Sensing. We have now carefully revised the manuscript in accordance with your instructions. Please find below, the specific changes made to the manuscript.

Please note that we attached a manuscript with track changes where you may assess the specific changes made to the manuscript. Please also note that our response to both reviewer’s comments are submitted to each reviewer, since some comments made by the two reviewers are interrelated. Hereby, we hope to provide both reviewers with a full overview of the changes made to the manuscript.

Yours,

Thomas Nord-Larsen

Reviewer #1

The authors have improved the manuscript substantially. The introduction is well written now. Also the explanation on the image processing is much clearer now.

Lines 298-305:

Even though selected grid cell size is 10x10m2 the original sensor resolution (Sentinel-2) is larger for some channels and in addition the sensor point distribution function typically is larger than the nominal pixel size thus gathering reflectance from even larger areas. Also, if median functions are applied to several images over same pixel, would it not be affected by the locational errors in between images and this  would add spatial variation to the output image.

Reviewer is right that accuracy of S-2 geo-referencing should be mentioned and kept in mind when doing such type of analysis. In the revised MS we added (l: 152-155):

After the activation of the Global Reference Image (GRI) geometric refinement, the geometric accuracy of S-2 multi-temporal co-registration is now better than 0.3 pixel (at 2σ confidence level) [48].

48. Gascon, Ferran, Catherine Bouzinac, Olivier Thépaut, Mathieu Jung, Benjamin Francesconi, Jérôme Louis, Vincent Lonjou, Bruno Lafrance, Stéphane Massera, Angélique Gaudel-Vacaresse, Florie Languille, Bahjat Alhammoud, Françoise Viallefont, Bringfried Pflug, Jakub Bieniarz, Sébastien Clerc, Laëtitia Pessiot, Thierry Trémas, Enrico Cadau, Roberto De Bonis, Claudia Isola, Philippe Martimort, and Valérie Fernandez. "Copernicus Sentinel-2a Calibration and Products Validation Status." Remote Sensing 9, no. 6 (2017): 584.

For the NFI plots, of course it is relevant to select good observations for the training data. 
But one could use all the plots for target data although the large size of the SSU plot sets limitations, e.g. plots partly on other land, and some compromises are needed. Compare this to the situation when a classified map should be validated by selecting a statistical sample of plots. In this case all the points or pixels within the forest map (the population) should have a probability to be selected to the sample.

In response to the reviwers comment, we included a description of the more common procedure of obtaining ground truth from e.g. photointerpretation with a finer resolution than the resulting map, which allows unique classification of each individual pixel. We further included references to extensive reviews, dealing with these matters (l: 239-247):

“The use of NFI sample plots for pixel-based classification of remotely sensed imagery may be problematic when labelling is obtained from the 706 m2 large circular sample plots, but the classification is subsequently trained and evaluated on 100 m2 (10x10 m) pixels [55]. In this case, the labelling obtained from the entire plot may not apply to the individual pixels encompassed (i.e. when one sample plot spans several land cover classes), leading to poorer training of the classifier. This is contrasting the common practice of selecting a statistical sample for classification or evaluation from an already classified map commonly derived from data with similar or finer resolution than the resulting map, where all pixels may be uniquely classified [56].”

55. Foody, Giles M. "Status of Land Cover Classification Accuracy Assessment." Remote Sensing of Environment 80, no. 1 (2002): 185-201.
56. Stehman, Stephen V., and Giles M. Foody. "Key Issues in Rigorous Accuracy Assessment of Land Cover Products." Remote Sensing of Environment 231 (2019): 111199.

 

Reviewer #2

The author improved manuscript accordingly. Conclusion Section still needs a major revision.

Please re-write it. No need repeat again the result in conclusion. Please mention your finding and provide some recommendation and limitation for this study.

In the revised manuscript we already provided recommendations for future mapping efforts, i.e. that the use of multi-temporal images yields robust estimates and that these may be improved by a finer temporal scale. However, in response to the comments also posed by reviewer #1, we included a reference to the limitation related to the use of NFI data for training and validation:

“Albeit sample plots with ambiguous land-use or tree species characteristics posed a specific challenge in relation to selecting samples for training and validation, the NFI data provided an unbiased sample covering the entire domain, controlled in the field, and readily available to the user.”

According to normal scientific practice, it should be possible to read the conclusion separate from the main document. Consequently, our results must be presented briefly and related to the scientific development, as was already done in the revised manuscript. We are thus reluctant to rewrite the conclusions, leaving out the results of our study as otherwise suggested by the reviewer, as we believe that this would impair the readability of the conclusion. We encourage the editor to provide directions as to whether we should leave out results entirely from the Conclusion-section.

Author Response File: Author Response.docx