Multi-Temporal Speckle Filtering of Polarimetric P-Band SAR Data over Dense Tropical Forests: Study Case in French Guiana for the BIOMASS Mission

Round 1

Reviewer 1 Report

This paper proposes an extension of multi-temporal filtering techniques to the case of multi-channel data.

My personal understanding of this filtering approach is that a prior about intensity (or covariance matrix) is simply retrieved by using a very large window, and this information is afterwards used to guide the estimation of intensity (or covariance matrix) at a finer resolution. Accordingly, whether this procedure brings some benefit is entirely depending on local heterogeneity features. As such, an experimental work appears to be well justified.

The problem with this paper is that it appears to have been written for hyper-specialized eyes only, Accordingly, I think much effort should be paid to make this paper clearer, easier to read, and more immediate to understand from the point of the applications, especially since it is openly intended for a specific application in forestry

I thus recommend a major revision of the paper making it easier to read and taking into consideration the following more specific comments.

 

Introduction

• Please describe a bit more in detail the scope of this work in the introduction. In this form, the reader arrives at figure 1 without knowing what the “REF” and “MCMT” filters do. Moreover, I think a brief description of what t0 is could help readability.
• The text reads that “Our purpose is to compare the two filtering methods”, which I understand are the MCMT filter and the REF filter. In this sense, a description of what the REF filter does would help (a lot).

Section 3.1

• The text would read much better if you specify before equation 1 that each element of the vector p is a multi-looked intensity – I guess some readers could miss this aspect and assume it’s just a single look.
• Please note that based on equations 1-4 one could argue that you estimate sigma based on the knowledge of sigma (since you need C to derive f). Probably, it would be better to write here that C is retrieved using a larger window, under assumption of local homogeneity…

Section 3.2. -  Extension to the multivariate case:

• I understand that by “channel” you mean now polarization channel, so that matrices Tk are either 2x2 or 3x3. It would be better to make this clear in the text to help readability. Please also note that in the previous paragraph the term channel appeared to be used in association with time.
• The derivation of the multi-channel filters needs to be explained in deeper details and with a more solid mathematical notation. In this form, the reader is either forced to repeat the demonstrations by him/herself, decide for an act of faith, or simply drop this paper for another one.

 Section 4:

• If understand correctly, the difference between MCMT and REF is simply that MCMT estimation is guided by prior knowledge derived using a larger window. I guess that deciding how larger the second window is has a big impact on final result. Can you comment on that?
• I dare anyone detect any difference between REF and MCMT based on figure 3. You should either use much higher-quality images or drop the comment that “Images filtered with the MCMT filter have a less grainy and smoother appearance while retaining a very pronounced texture”
• It is hard to appreciate any advantage of MCMT over REF based on figure 4. Wouldn’t it be better to asses the impact on forest biomass estimation (at the end, this is what t0 is used for…).
• What do you mean by: “The NCI filter impacts the average PolSAR intensities in an equivalent way to the REF filter. The average is not a good criterion to differentiate between these two types of filtering”? I couldn’t find the definition of NCI filter in the text

 

Section 4.2.2:

• what is NCI?
• Why should the dispersion of Leq increase with increasing number of looks?
• You state that “ENL is not a significant criterion for studying the contribution of the MCMT filter applied to PolSAR data since it does not take into account the correlation of multi-channel and multi-polarised data”. This is confusing. If it is not significant, why include this section at all?

 

Section 4.2.3: based on figure 7, the REF estimator produces finer resolution. Accordingly, the reader is left with the doubt that the gain of the modified filter is mostly associated with the use of a larger window. Please comment on this point.

 

 

Author Response

Multi-temporal speckle filtering of Polarimetric P-band SAR data over dense tropical forests in French Guiana: application to the BIOMASS mission

Answers to reviewer 1

 

Comments and Suggestions for Authors

This paper proposes an extension of multi-temporal filtering techniques to the case of multi-channel data.

My personal understanding of this filtering approach is that a prior about intensity (or covariance matrix) is simply retrieved by using a very large window, and this information is afterwards used to guide the estimation of intensity (or covariance matrix) at a finer resolution. Accordingly, whether this procedure brings some benefit is entirely depending on local heterogeneity features. As such, an experimental work appears to be well justified.

The problem with this paper is that it appears to have been written for hyper-specialized eyes only, Accordingly, I think much effort should be paid to make this paper clearer, easier to read, and more immediate to understand from the point of the applications, especially since it is openly intended for a specific application in forestry

I thus recommend a major revision of the paper making it easier to read and taking into consideration the following more specific comments.

We are really grateful for this very constructive review, especially for this interpretation which better puts the filtering approach in perspective. This comment about the regional estimate as prior has been added in the revised version (cf. section 3). We also agree on the lack of details in the original version, and we did our best to improve it with clarifications, as detailed in the following point-to-point answers.

Introduction

• Please describe a bit more in detail the scope of this work in the introduction. In this form, the reader arrives at figure 1 without knowing what the “REF” and “MCMT” filters do. Moreover, I think a brief description of what t0 is could help readability.

We agree, this has been explained more clearly in the introduction.

• The text reads that “Our purpose is to compare the two filtering methods”, which I understand are the MCMT filter and the REF filter. In this sense, a description of what the REF filter does would help (a lot).

We agree, some terms are not detailed enough, to improve this lack, introduction has been improved. In particular, the REF and MCMT filters are detailed before explaining the means used to compare them.

Section 3.1

• The text would read much better if you specify before equation 1 that each element of the vector p is a multi-looked intensity – I guess some readers could miss this aspect and assume it’s just a single look.

Indeed, the previous derivations were not detailed enough to avoid this misunderstanding. To this end, lines preceding equation (1) have been completed. 

• Please note that based on equations 1-4 one could argue that you estimate sigma based on the knowledge of sigma (since you need C to derive f). Probably, it would be better to write here that C is retrieved using a larger window, under assumption of local homogeneity…

It is indeed a relevant way to present the filter, thank you for this suggestion. We thus added this note in the revised version (cf. section 3).

Section 3.2. -  Extension to the multivariate case:

• I understand that by “channel” you mean now polarization channel, so that matrices Tk are either 2x2 or 3x3. It would be better to make this clear in the text to help readability. Please also note that in the previous paragraph the term channel appeared to be used in association with time.

We agree that these terms can be confusing, so to simplify understanding, the term "multi-channel" relating to temporal variations has been replaced in section 3.2 by the term "multi-date". however, the term "multi-channel" for polarimetric variations has been retained.

• The derivation of the multi-channel filters needs to be explained in deeper details and with a more solid mathematical notation. In this form, the reader is either forced to repeat the demonstrations by him/herself, decide for an act of faith, or simply drop this paper for another one.

We agree that some steps were presented to shortly. Moreover, we had omitted to correct some typing errors. Equation (7) has been corrected and the explanations relating to this section have been completed.

Section 4 :

• If understand correctly, the difference between MCMT and REF is simply that MCMT estimation is guided by prior knowledge derived using a larger window. I guess that deciding how larger the second window is has a big impact on final result. Can you comment on that ?

We agree, the choice of the multi-look window size defined for the calculation of  was not mentioned explicitly. It is defined as twice the size of the multi-look window used for . It is indeed a relevant point, we choose in this study a regional multi-looking twice stronger than the local one. This choice results from an empirical trade-off between a regional multilooking large enough to serve as prior free of speckle, and small enough to deal with an homogeneous region. This has been specified in section 4.1 of the revised version.

• I dare anyone detect any difference between REF and MCMT based on figure 3. You should either use much higher-quality images or drop the comment that “Images filtered with the MCMT filter have a less grainy and smoother appearance while retaining a very pronounced texture”

We agree, the differences between REF and MCMT are not visually significant. The comments related to the observation in figure 3 have been modified accordingly.

It is hard to appreciate any advantage of MCMT over REF based on figure 4. Wouldn’t it be better to assess the impact on forest biomass estimation (at the end, this is what t0 is used for…)?

Indeed, this part was a little confusing and the usefulness of the different indicators was not clear enough. With this in mind, section 4 has been reorganized and reworked. The importance of conserving the average intensity during speckle filtering is thus highlighted, as well as the gain in resolution observed for MCMT filtering. In particular, a figure has been added to compare more precisely the polarimetric  indicators resulting from the two types of filtering.

• What do you mean by: “The NCI filter impacts the average PolSAR intensities in an equivalent way to the REF filter. The average is not a good criterion to differentiate between these two types of filtering”? I couldn’t find the definition of NCI filter in the text

We apologize, this is a naming error related to an earlier version of the article. The acronym NCI stands for MCMT and has been replaced in the text.

Section 4.2.2 :

• what is NCI?

Again, our apologies (see previous comment). The necessary corrections have been done.

• Why should the dispersion of Leq increase with increasing number of looks?

It is indeed not expected under the common statistical assumptions (independent and identically distributed realizations between each pixel). Actually, this result illustrates very well the impact of pixel correlation within specific regions of the image, especially between ROIs 1 and 2: both are covered by a dense vegetation but ROI 1 is characterized by a significant terrain topography which generates a higher level of inter-correlation between pixels, hence a degraded efficiency of the multilook (and smaller ENL). This comment has been also added in the revised version.

 

• You state that “ENL is not a significant criterion for studying the contribution of the MCMT filter applied to PolSAR data since it does not take into account the correlation of multi-channel and multi-polarised data”. This is confusing. If it is not significant, why include this section at all?

The aforementioned ENL formula based on the Gaussian hypothesis is a common indicator for estimating the performance of speckle filters. For its application to PolSAR P-band data specific to the observation of dense forests, it is important to show its limitations to characterize the performance of the MCMT filter. Moreover, these limitations highlight the importance of developing ad hoc indicators to study the performance of speckle filters for P-band data. To this end, section 4.2.2 has been completed.

Section 4.2.3:

• Based on figure 7, the REF estimator produces finer resolution. Accordingly, the reader is left with the doubt that the gain of the modified filter is mostly associated with the use of a larger window. Please comment on this point.

It is true that with the increase in multi-look, the added-value of the MCMT filter is lower. Nevertheless, with a multi-look window size equivalent to 50 m, the details visible in figure 7 for the  REF correspond to noise. This is confirmed by the histograms in figure 8, where a greater dispersion of slopes around the diagonals is observed for the results from the REF filter. This analysis was not very developed in the first version of the article, we apologize for this. For this new version, the analyses of images 6, 7 and 8 have been completed.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

This manuscript presents an extension of an intensity based multi-temporal speckle filter to a multi-channel multi-temporal SAR speckle filter. The main idea behind this paper is the extension of a well known multi-temporal filter to a multi-channel multi-temporal form. Moreover, the paper compares a DEM derived slope in the azimuth direction with the azimuthal slope derived from the filtered PolSAR coherency matrix to evaluate the performance of the proposed extension. They also used the average intensity in addition to the equivalent number of looks to evaluate the filtering performance of the proposed extension.

The manuscript is well written. I also think the topic is interesting and the methodology and results appear sound.

However, the novelty of the proposed work is not clear. The proposed extension from single channel multi-temporal SAR image to a multi-channel multi-temporal SAR image is straightforward. The quality measurement metric discussed in the paper is also known in PolInSAR literature. Therefore, I suggest to the authors to consider boosting the novelty of the manuscript, possibly by adding an analysis from additional scene different from the one shown in the manuscript.

Therefore, I wouldn’t recommend the paper for publication in its present form, but I would strongly encourage the authors to improve the manuscript.

Author Response

Multi-temporal speckle filtering of Polarimetric P-band SAR data over dense tropical forests in French Guiana: application to the BIOMASS mission

Answers to reviewer 3

This manuscript presents an extension of an intensity based multi-temporal speckle filter to a multi- channel multi-temporal SAR speckle filter. The main idea behind this paper is the extension of a well known multi-temporal filter to a multi-channel multi-temporal form. Moreover, the paper compares a DEM derived slope in the azimuth direction with the azimuthal slope derived from the filtered PolSAR coherency matrix to evaluate the performance of the proposed extension. They also used the average intensity in addition to the equivalent number of looks to evaluate the filtering performance of the proposed extension. The manuscript is well written. I also think the topic is interesting and the methodology and results appear sound. However, the novelty of the proposed work is not clear. The proposed extension from single channel multi-temporal SAR image to a multi-channel multi-temporal SAR image is straightforward. The quality measurement metric discussed in the paper is also known in PolInSAR literature. Therefore, I suggest to the authors to consider boosting the novelty of the manuscript, possibly by adding an analysis from additional scene different from the one shown in the manuscript. Therefore, I wouldn’t recommend the paper for publication in its present form, but I would strongly encourage the authors to improve the manuscript.

Authors' reply:

We are grateful for this positive overall feedback, and we do understand the reviewer's concerns about the novelty of this contribution given that the main insights haven't been sufficiently emphasized to be fully caught.

 As stated in the original abstract, the purpose of this paper is twofold, the first insight being focused on the filter extension to multi-temporal and multi-channel SLC data, the second on the correlation between the POA and the azimuthal slopes as a metric to assess the filtering performance.

For the first point, we must have better highlighted the SLC feature, since most filtering approaches are restricted to SAR intensities. In our case, the kernel of our filter cannot be fully optimized with the phase information embedded in the SLC data, but can still be applied to fully polarimetric cases, providing a solution to feed any kind of process (such as POA estimation) based on SLC data. This explanation has been added to the revised version and should refocus the reader attention (see section 4.3).

Regarding the second point, we agree about the common knowledge of the polarimetric SAR data capabilities to retrieve azimuthal slopes, but it consists in using the correlation between the POA and an independent ground slope estimation as a filtering performance indicator, in order to overcome the limitation of the ENL which is widely applied in spite of its major limitations in the case of forested land-cover which might introduce a high level of auto-correlation within the images, and thereby the invalidity of the ENL formula stemming from the Gaussian hypothesis. Although this behavior might be empirically known by expert users, the related literature is really rare so that we are still convinced that our contribution will be relevant for the community, as rewritten in the revised version (see section 4.3.1).

Besides, we can also acknowledge that the above-mentioned two points could have been submitted separately, but merging them is relevant in the up-to-date framework of the BIOMASS mission, given the expected temporal sequences of polarimetric SLC data and the use of P-band which overcomes the signal penetration issue to derive ground slope information.

Last, considering the database extension would be very interesting, but we are facing the lack of suitable P-band SAR data, especially with acquisition scenarios close to the future BIOMASS ones with at least several days between passes. This explanation has been also made more explicit in the revised version.

Reviewer 3 Report

P-band synthetic aperture radar is a remote sensing technique using low frequency microwave. P band waves can penetrate through dense media such as forest cover, where they interact with a variety of scatterers like forest canopy, branches, trunks and even the ground. Therefore, P band radar has attracted growing interests. This paper developed an ad-hoc performance indicator to better estimate the resulting speckle reduction in the case of dense vegetation. The idea is novel and the paper is well constructed. Before possible publication, the following comments can be considered.

 

In section 3, the Multi temporal and multi channel speckle filter is not described clearly.

1) Expression (5) is not described clearly, the author could discuss more on (5) to tell how (5) will be employed in the filtering or in the next sub-section.

2) The authors should discuss more on expression (9). How (9) will perform in the process. 

 

It is recommended that the author could rewrite section 3 for better presentation.

Author Response

Multi-temporal speckle filtering of Polarimetric P-band SAR data over dense tropical forests in French Guiana: application to the BIOMASS mission

Answers to reviewer 3

P-band synthetic aperture radar is a remote sensing technique using low frequency microwave. P band waves can penetrate through dense media such as forest cover, where they interact with a variety of scatterers like forest canopy, branches, trunks and even the ground. Therefore, P band radar has attracted growing interests. This paper developed an ad-hoc performance indicator to better estimate the resulting speckle reduction in the case of dense vegetation. The idea is novel and the paper is well constructed. Before possible publication, the following comments can be considered.

In section 3

• The Multi temporal and multi-channel speckle filter is not described clearly.

We apologize for this lack of clarity, in agreement with the requests of reviewer 1. We have substantiated our explanations and added equations to make the filtering method easier to understand.

• Expression (5) is not described clearly; the author could discuss more on (5) to tell how (5) will be employed in the filtering or in the next sub-section.

Indeed, the explanation of the assumptions of data distribution and correlation was too rapid. To clarify the filtering process, we have detailed equation (5) through 4 equations, specifying in particular when the assumptions concerning the filtered data are applied. 

• The authors should discuss more on expression (9). How (9) will perform in the process. 

To support the filtering proposed in section 3.2.2, we have added the reference [15] to explain more clearly how the MCMT filter applies in the polarimetric case, and we have completed the explanations related to equation (9)

• It is recommended that the author could rewrite section 3 for better presentation.

We agree with those requirements. It has been well taken into account, section 3 has been rearranged and completed for more clarity and for a better understanding of the filtering method proposed in this article.