|Identification of blowing snow particles in images from a multi-angle snowflake camera |
Mathieu Schaer, Christophe Praz, and Alexis Berne
The authors present a novel analysis of two precipitation datasets that indicates a potential data-driven method for identifying falling precipitation from blowing snow precipitation. Differentiating water in these two stages of the snowfall life-cycle is critical for having an accurate water balance for a given region and all hydrological and climatic conclusions that come from that water balance. The authors rely on a Multi-Angle Snowflake camera and a Gaussian mixture model to identify two end-states of the blowing-snow to fresh-snowfall spectrum. The strength of this probabilistic approach is that one can define the mixture model given site-specific data and, through a normalized angle metric, the probability that a given image corresponds to one of the two end-states.
The technical nature of this paper is quite high, and as mentioned by previous reviewers, the audience with sufficient statistical background to entirely follow the manuscript is likely to be a small subset of the cryosphere community. I have one foundational criticism of the results, one criticism of the presentation, and several minor comments below.
Primarily, it is as of yet unclear to me whether or not this Gaussian mixture method is necessary or beneficial for categorizing images. Figs 6 and 7, and to some point Figure 14 seem to suggest that the dominate separating factor for the two end-states is the image frequency. Though there is a low finite upper limit on this frequency, this is a loose proxy for snow particle flux in much the same way as a particle counter gives you an index of how many times a sensor is triggered. Physically, it makes sense for the snow flux to differentiate these two regimes as the settling velocity of falling particles is much lower than potential transport speeds, and the potential rates of snow transport by the two methods are quite different. That being said, it would benefit the manuscript greatly if the authors could show that all the technical machinery of the Guassian mixture model and the addition of the other image analysis metrics (Distance Transform, Squared fractal index, Dmax) are indeed necessary to have accuracy at this order of magnitude and that they are not superfluous technical additions. Other advantages that I may have overlooked would also benefit from being highlighted more.
I think it would be illuminating for a broader audience and enhance the transparency of the manuscript if it was clearly acknowledged that the normalized angle does not actually give any indication of what proportion of a given image is blowing snow versus precipitation, but actually only indicates what the probability is that an image is one of the two end states according to their training data. As the methods are currently described, this is my understanding of the Gaussian mixture model output. If this is inaccurate, it would also be of benefit to rectify future misunderstandings with further clarification. Furthermore, for technical the paper is, the validation appears to be largely qualitative, with a tendency towards arguing “typically there is more blowing snow here than there…”
The flow of the paper is currently a limiting factor in its readability. Ideas are not introduced or referenced in a natural order. Some comments along these lines are discussed below. The English needs improving.
Specific comments are as follows:
P2L4-10: This drifting versus blowing snow designation is unnecessary, and the authors are inconsistent in the use of it. The more technical modes of creep, saltation, and suspensions would be more appropriate differentiations.
P2L18-20: Please refrain from saying obviously as it undervalues the work.
P2L27: What motion detector system? This has not be referenced yet.
L28-29: How was this adapted, because Praz et al., 2017 says nothing about “blowing snow”, “drifting snow”, or “fragmented grains”.
P2L31-33 Unclear motivating statement
P3L17 Does this mean the cameras are not synchronously taking pictures? If so, the sampling frequency is 1 Hz, a distinction of great relevance for blowing snow measurements, where counts scale with flux. This is confusing for Figure 3. What rate is maximal?
P3L20: A better comprehensive reference of (blowing) snow measurement techniques is Kinar and Pomeroy (2015).
P4L5-13 Refer to the Table, and use the actual months that contained observations (8 days Nov-Jan, etc.), so as to not overstate the amount of data used. Consistently reference the dates (i.e. not just years for one dataset and years and months for the other).
P4L13 Was this 11.5 days total? Please clarify.
P4L15 Choose not chose.
P4L15 Rephrase “enough”. A sufficient number of?
P4L16 classes not class.
P4L16 Especially? How so?
P4L16 Rephrase “appeared less trivial than expected”
P4L17 For those that study the cryosphere, but not East Antarctica, how far away are these stations, and are they similar?
P4L19 “For the sake of generalization…” is not a sentence
P4L20 Correct the phrase “hydrometeors types as well as snowfall rate”
P5L5 The sentence beginning with “Similarly” seems like an incomplete or unconnected thought
P5L7-9 Why would the image frequency need to be lower than the median during pure precipitation? Is there a physical basis for that?
P5L14-16 Please be consistent with tenses throughout the paper “we noticed….one could notice”
P5L16 Is augment the right word choice here?
P5L19-20 What is this “uncertainty?” Is this 4263 unique instances, or 1421 unique timesteps? Previously commented.
P7L6-7 Can you make a mention of the focal length of these cameras? i.e. are all particles always in focus? This is critical for distinguishing blowing snow particles from falling snow.
P7L12 What is the window size of your median filter? Median filters have an effective smoothing, depending on window size, essentially blurring all edges.
P7L16 “rarely”, not “hardly”. “in” not “on”. get rid of few or replace with multiple.
P7L23 Rephrase “by a too long period of time”
P7L27 A more standard and original reference to cite for decomposition of blowing snow grains is Schmidt, 1980 “Threshold wind-speeds and elastic impact in snow transport”
P7L29 A bit more effort should be made to cite papers where these ideas originated (Budd et al., 1966 “The byrd snow drift project: outline and basic results” and Budd 1966 “The drifting of nonuniform snow particles”)
P8L1-3 What does that mean? Sentence starting “As…”
P9L3 Replace pertinent
P10L7 This choice seems awfully arbitrary. Can it be backed up by anything?
P10L14-15 And what is the significance of having the largest S value?
P10L17 Refers back to my original concern. How do we know all the other machinery surrounding the image frequency is necessary?
P11L8 You mean many-fold not manifold
P12L1 Clear it up and define what the vectors are: “x=(image freq, …)”
P12L2 “for this purpose”
P12L21 Rephrase “In words”
P13L6-7 Where is the degree of mixing actually verified? I only see uncertainty later on, not something physical. Refer to second major comment.
P13L7-8 This “likelihood” is entirely contingent on your method working. If it does not work, being near the decision boundary means inconclusive.
P13L11-13 Rephrase sentence beginning with “Nevertheless…”
P15L1 Rephrase “The terms have usually opposite signs…”
P16L5-7 Again, how was this generated?
P16L12-13 This is not immediately clear. Precipitation particles are most clearly evident in the top subplot, whereas the blowing snow (combined with mixed?) are dominant below. Please clarify.
Fig10 Why is there a peak around 45 degrees? Does this not imply some tendency towards inconclusive results as the probability is neither one nor the other? I do not recall a training specifically for mixed grains.
P17L5 Reword sentence beginning with “The proposed…”
Fig 11 Are these results? What ground truth do we have for a comparison? These results seem largely qualitative.
P18L1-2 This seems to imply that you are classifying each camera’s images separately. How often did the three cameras agree or disagree in the same few hundredths of a second? This comparison would help bolster the claims that the authors are coming up with self-consistent results. Furthermore, the image classifications self-consistent if one was to remove image frequency? That is, do the other image analysis metrics represent something actually physically relevant, or are the results then “noisy.”
P18L3-4 Rephrase “the type…and mixed”
P18L5-6 Rephrase the sentence beginning “Finally…”
Fig 12 Really hard to interpret. Please use a binary image or increase the contrast.
Fig 13 Interesting! And what’s the actual truth here? How likely is it that there was “Pure Blowing Snow” happening concurrently with “Pure Precip”?
Fig 14 Use more obvious overlapping patterns. It is unclear what is happening when more than two distributions start overlapping
P20L10 Where is Davos blowing snow? If this is blown right off of the fence tops, it should look something like a mix of fresh precip and blowing snow as it has not had a chance to fragment on the ground.
P21L15-16 This has not been convincingly argued.
P21L18-19 This is in effect a methods paper with minimal validation. At the moment, these conclusions are suspect.
P22L1-3 Why not use the actual weather station data nearby, instead of relying on statistics from other years. This reliance makes the conclusions weaker than necessary.
P22L19-20 This should be mentioned much earlier, as there is no reason this assumption should hold.
Kinar, N. J., and J. W. Pomeroy (2015), Measurement of the physical properties of the snowpack, Rev. Geophys., 53.