Open-source algorithm for detecting sea ice surface features in high-resolution optical imagery

Wright, Nicholas C.; Polashenski, Chris M.

doi:https://doi.org/10.5194/tc-12-1307-2018

Articles | Volume 12, issue 4

https://doi.org/10.5194/tc-12-1307-2018

© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/tc-12-1307-2018

© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 12, issue 4

Research article

|

12 Apr 2018

Research article |

| 12 Apr 2018

Open-source algorithm for detecting sea ice surface features in high-resolution optical imagery

Nicholas C. Wright and Chris M. Polashenski

Download

Final revised paper (published on 12 Apr 2018)
Preprint (discussion started on 18 Sep 2017)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review of Wright and Polashenski', Anonymous Referee #1, 21 Oct 2017
- AC2: 'Response to comments of referee #1', Nicholas Wright, 18 Dec 2017
RC2: 'Review of "Open Source Algorithm for Detecting Sea Ice Surface Features in High Resolution Optical Imagery"', Anonymous Referee #2, 21 Oct 2017
- AC3: 'Response to comments of referee #2', Nicholas Wright, 18 Dec 2017
RC3: 'An interesting application for Polar program', Anonymous Referee #3, 04 Nov 2017
- AC4: 'Response to comments from referee #3', Nicholas Wright, 18 Dec 2017
AC1: 'Revised manuscript with mark-up', Nicholas Wright, 18 Dec 2017

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

AR by Nicholas Wright on behalf of the Authors (29 Dec 2017) Author's response Manuscript

ED: Reconsider after major revisions (15 Jan 2018) by Bert Wouters

ED: Referee Nomination & Report Request started (19 Jan 2018) by Bert Wouters

RR by Anonymous Referee #1 (19 Jan 2018)

Suggestions for revision or reasons for rejection

With regards to estimate the required sample size using equation in line 568, I am not sure their calculation for required number of samples is correct in the current revision - or at least that they have used incorrect values. This needs to be addressed before publication as I think it alters the size of sample required for analysis.

I still think this paper should be published but the calculations need to be correct. Two things need to be done: 1) the calculations and values used need to be checked to make sure they are correct. 2) They need to state what they mean by margin of error (either % error of best estimate or absolute error) and the absolute error needs to be used in the calculation of sample size.

My specific comments about the problems I see are here.

To find an acceptable sample size they use:

n = (z * sigma / margin_of_error)**2 (1)

where n is the number of samples required to get margin_of_error, z is the z-score (x_est - X/sigma), sigma is the standard deviation and margin of error is, in this case, the accepted error of the measurement. They want to find n such that the margin of error is 5% and appear to use a value of 0.05 in equation 1.

Based on the text you sent, it appears they use z=1.96, sigma=0.2 and margin_of_error=0.05. Using these values and their equation (equation 1 here), I get n~62, not n=64 given in the paper. If I set z=2 (e.g. a 2*sigma margin of error), I get n=64. They either need to revise n or change z to 2, in which case the confidence interval is 95.4%. This is not a big deal (using z=1.96 versus z=2.0 does not change the sample size much), but it is frustrating if you try to follow their calculations. The resulting sample size should be consistent with the parameters for the equation.

A bigger issue is how they use margin of error. They say they want to find a sample size that will give a 5% margin of error with ~95% confidence interval. I take a margin of error of 5% to mean 5% of the best estimate of melt pond (or ice) fraction. Based on Figure 14a, the best estimate of pond fraction for the image they use is ~0.4. The corresponding margin of error (based on my understanding) is 0.02 (0.05*0.4). Using equation 1, this gives a sample size of almost 400, not the ~60 they give in the paper. A margin of error of 0.05 in terms of fraction is 12.5%. This means that more than 6 times more samples are required to get the desired margin of error of 5%. I give my reasoning for this below.

Reasoning
---------------
Equation 1 follows from the statement of uncertainty of a measurement

x_est = x_best +/- margin_of_error (2)

where x_best is the mean of n measurements. margin_of_error can be given in the same units as x_best (here a dimensionless fraction [0,1]) or a percentage. The percentage is 100*margin_of_error/x_best. margin_of_error is given by

margin_of_error = t*standard_error = t*sigma/sqrt(n).

t is the number of standard errors a fraction P other measurements are expected to fall from x_best. In their case, P is 95%, which corresponds to a t of 1.96. This is equivalent to a z-score of 1.96. margin_of_error, standard_error, and sigma are all in the same units as x_best.

Hide

RR by Anonymous Referee #2 (07 Feb 2018)

Suggestions for revision or reasons for rejection

The revised manuscript has greatly improved, and I thank the authors for giving careful attention to addressing the raised concerns. The presented work makes a nice contribution to the field, and it has been a pleasure reviewing it. I have remaining concerns that should be addressed before final publication:

Line 55. "but were historically limited" High resolution imagery is still spatially limited in basin-wide coverage. It's incorrect to suggest that it's no longer an issue.

Line 66. "none have been challenged by imagery collected across the seasonal evolution of the ice" Fetterer and Untersteiner, 1998 studied the seasonal evolution of the ice surface from a Eulerian perspective, much like this analysis. Rosel et al., 2012 studied the seasonal evolution of the ice surface from a basin-wide perspective. Perovich et al., 2002, Arntsen et al., 2015, and Webster et al., 2015 studied the seasonal evolution of the ice surface from a Lagrangian perspective. It's not clear why the authors are suggesting that this hasn't been done before with airborne and satellite imagery.

Lines 260-262. Is it correct that, once melt onset occurs, shadows are not accounted for? Please make this explicit, as it will affect the interpretation of melt pond retrievals during early pond formation.

Lines 262-265. The approach in Webster et al. 2015 does not directly mask shadows as a separate class, but groups them with the ice/snow category.

Lines 332-335. More details are needed here. The explanations should be more explicit, "We use image acquisition dates as a proxy for melt onset." Which dates were used as thresholds for defining melt onset, and is there a latitudinal dependency for these dates?

Hide

ED: Publish subject to minor revisions (review by editor) (12 Feb 2018) by Bert Wouters

AR by Nicholas Wright on behalf of the Authors (22 Feb 2018) Author's response Manuscript

ED: Publish as is (11 Mar 2018) by Bert Wouters

AR by Nicholas Wright on behalf of the Authors (21 Mar 2018) Manuscript

Short summary

Satellites, planes, and drones capture thousands of images of the Arctic sea ice cover each year. However, few methods exist to reliably and automatically process these images for scientifically usable information. In this paper, we take the next step towards a community standard for analyzing these images by presenting an open-source platform able to accurately classify sea ice imagery into several important surface types.