Review of: Evaluating Snow Depth Measurements from Ground-Penetrating Radar and Airborne Lidar in Boreal Forest and Tundra Environments during the NASA SnowEx 2023 Campaign

This is a nice tidy paper with useful data, but I think it reaches conclusions that a more thorough analysis might contradict. I spent a while studying figures S3, S4 and S5, the real heart of the paper, and think it would improve the paper if the authors went back to these figures and spent a bit more time thinking about them and how the various data traces relate to each other.  

First off, Figures S3, S4, and S5 are the real results and ought to appear in the paper itself, not just as supplemental material. The statistical summaries in Figure 2 are useful, but they don’t allow a reader to examine the data in detail.  Moreover, the data in Figure 2 are from the first two moments in statistics (mean and variance), but these do not make use of the spatial organization of the snow depth data (e.g., the relative relationships of ups and downs in a depth transect). Figures S3, S4 and S5 do that, and provide a reader with better sense of how well the remote sensing sampling (GPR and lidar) reproduced the depth point data profile. I would highly recommend to the authors reading a paper by Blöschl and Sivapalan (1995) that discusses support, extent and spacing in snow sampling. In this study, where those differ so much, it might help in framing the comparison and conclusions.

Blöschl, Günter, and Murugesu Sivapalan. "Scale issues in hydrological modelling: a review." Hydrological processes 9, no. 3‐4 (1995): 251-290.

Here are a few points that are not addressed in the paper, but could be important:

1. The GPR surveys were done at very small scale (a few meters), so despite the GPR’s larger areal footprint, we might expect the point-to-point co-registration between the GPR data and the excavated snow depths to be excellent. The co-registration between the airborne lidar and excavated depths is done by GPS, and has stated potential error of 50% of the transect length. The lidar requires two co-registrations: between snow-on and snow-off acquisitions (which gets done months later), and between that result and the excavated depths. The lidar aircraft flies at least 2000’away, and all its depths, and position of the depths, relies on a black-box solution using Metashape or Pix4D. I don’t think co-registration of lidar to the depth transects can be assumed, and perhaps it is prudent to assume it is likely to be off.
2. The authors ascribe the offsets between the GPR and ruler depths to air gaps, but say nothing about what happens when one compares depths averaged over a substantial circular area (over a bumpy substrate of tussocks) to a two-dimensional profile in which the edges of 3D bumps appear as a simple 2D curve. This support (Blöschl and Sivapalan, 1995) mismatch could easily explain some or even a lot of the differences between the GPR and the ruler depths. Ironically, though the lidar is run thousands of feet away from the target, its support has the putative dimensions of a photon (or there-abouts) and so theoretically matches the excavated depth support better.
3. Which brings me what I call the differences in work flow of the two remote sensing methods (e.g., the processing that has to happen with these two types of remote sensing efforts to produce a depth value.) For the GPR, a human picks the ground surface, a little subjective, but an expert process, and done for a very limited area. For the lidar, that happens in a black box, requires two flights separated in time, and any co-registration mismatch between the first and second acquisitions will produce errors (not biases).  It is done for a huge area all at once. So I would question how relevant it is to test airborne lidar data that can cover many square kilometers using such limited scale (5- to 10-m) ground trench data.
4. That said, looking at Figures S3 to S5, the lidar profiles seem to occur in two modes for the taiga areas: either the lidar falls right on top of the excavated depth profiles (FLCF 11 March DN091) or it is way off (FLCF 9 March WB032), but in many cases, parallel to the excavated depth profile. This binary behavior suggests to me that in the lidar data we are seeing both biases and errors. The biases are where something like frost heave or void spaces have produced differences between the lidar and the probe depth; the errors are where the co-registration or the GPS bundle solution, or something else, has gone wrong.
5. The Kuparuk-Toolik results also seem to show a similar behavior: there are some nice lidar to depth matches but also some poor ones:  for example, 9 Mar N789 is “on” for both the GPR and the lidar, but for 11Mar A739 the lidar has a distinct low bias across the whole transect. That stands as an argument that substrate character alone is not necessarily the factor driving the bias in the lidar.
6. Which brings me to the puzzling results from the Arctic Coastal Plain (ACP). I have worked extensively with Chris Larsen collecting airborne lidar across the ACP, and I have validated those results with literally thousands of on-the-ground snow depths (see Sturm et al., 2019 (this report to BLM and the USFWS is hard to find now, so I attach it); see also Nolan et al., 2015). It has been our experience in this work that often there is an affine translation in the lidar raster depth surveys that needs to be corrected to produce a useful survey. Basically, due vagaries in the 3D GPS solution, or perhaps it is minor issues with the GPS constellation, the entire survey may float too high or drop too low. Anyway, that affine translation needs to be corrected by using ground validation data. It doesn’t take much ground data to do so. If there are no in situ snow measurements, we have used snow-free patches to do this. So the lidar data for the ACP 13 Mar A522, adjusted downward about 12 cm, gives a fine match to the excavated depth data, suggesting that some of the lidar data have issues like this.  Apropos to the SnowEx goal of validating an airborne radar retrieval, I would simply take the extensive magnaprobe field data from SnowEx, and “calibrate” the lidar flight data against it to produce a B-level corrected product, then use that for comparison to the radar results.

Nolan, M., Larsen, C., & Sturm, M. (2015). Mapping snow depth from manned aircraft on landscape scales at centimeter resolution using structure-from-motion photogrammetry. The Cryosphere, 9(4), 1445-1463.

Summary:  In the end, I felt like this paper was comparing apples to oranges, or perhaps cherries (GPR) to watermelons (lidar), and of course, finding differences.  The lidar is designed to cover hundreds of square kilometers, but at the cost of a complex technical solution that can show accuracy drift and needs to be “calibrated” in practical use. The GPR is good at providing a detailed solution on a small patch of snow, using a human interface for handling the complexity of “picking” the base of the snow. It imposes strong spatial averaging.  The data presented are interesting data, but I think the explanations for the differences in remote sensing methods presented in the paper are neither fully correct nor nuanced enough. I suggest going back after reading Blöschl and Sivapalan (1995) and looking at those supplemental figures more closely.

Minor Points:

Lines 26-28:  Snow also strongly impacts the ecology of these regions: snow influences caribou winter range selection (Pedersen et al., 2021) and vegetation phenology (Kelsey et al., 2021), and provides winter refuges for a diverse range of animals (Penczykowski et al., 2017). 

This is a point I try to make (usually unsuccessfully) to young investigators: it is silly to reference a 2017 or 2021 reference to make general points about snow in the North. There are so many earlier papers that established that point…some dating back 70 years or more.  To fail to credit all that great older seminal work is to suggest it didn’t happen. I think I would prefer no citations to buttress the statement in the text than a cursory sprinkling in of some newer citations that suggest the old work never happened.  Similarly, where the authors cited my work (Sturm and Liston, 2021) as evidence for wind slabs on the tundra and faceted grains in the taiga, I cringed. That 2021 paper is global in scope; to lead a reader to a useful reference on depth hoar of wind slab, how about citing my mentor, Carl Benson (now 98):

Benson, C.S. 1967:  Polar Regions Snow Cover, In Physics of Snow and Ice : Proceedings, 1(2), 1039-1063. International Conference on Low Temperature Science. I. Conference on Physics of Snow and Ice, II. Conference on Cryobiology. (August, 14-19, 1966, Sapporo, Japan),

or if you must:

Benson, C. S., & Sturm, M. (1993). Structure and wind transport of seasonal snow on the Arctic slope of Alaska. Annals of Glaciology, 18, 261-267.

Line 74:  There are tussocks, and there are hummocks, and there are ice wedges and polygons.  All combine to make the tundra a bumpy surface. Perhaps be a bit more general here.

Line 141: Can we assume that after the GPR pass there were foot holes and sled marks in the snow, and that the excavation resulted in both a trench and heaped up pile of snow behind the trench?  So there were many square meters of messed up snow. Perhaps any lidar that was done post-excavation ought to be culled from the paper.

Line 155-Figure 2:  Pretty clear that the ACP and UKT data differ in some fundamental way for the lidar.  This then drives the difference between lidar and GPR for the tundra.  I think it might be more useful when presenting the taiga results to color-code the data by site rather than canopy/vegetation height. Then we could see if there is a site bias in the lidar for the taiga as well as the tundra.

Line 217-219:  Conclusions: Sorry, but I am just not convinced this statement is valid. The data indeed differed between methods, but of there are so many variables affecting the results of each remote sensing method that citing just void spaces and vague reference to vegetation effects gives the wrong impression. Don’t ignore the vast differences in support, spatial sampling between the methods, as well as the work-flow differences in the data reduction.  

Dear authors, I had the opportunity to review your manuscript entitled "Brief Communication: Evaluating Snow Depth Measurements from Ground-Penetrating Radar and Airborne Lidar in Boreal Forest and Tundra Environments during the NASA SnowEx 2023 Campaign".

General comments:

In your work, you assess the lidar accuracy to map snowpacks in high-latitude environments, with a focus on boreal forests and tundra environments. You perform a comparison with GPR and manual excavation in several transects. The manuscript is clear and well-written, and highlights the importance of assessing instrument uncertainties in mapping the snow accumulation.

I am not a lidar expert; therefore, I do not comment on it. However, I performed GPR measurements in snowpacks. The GPR data collection, the instruments used, and the resulting radargrams are of good quality.

I appreciated your work because you highlighted the possible causes of the observed uncertainties, which I find very useful for further research. You do not always investigate in detail such causes, which you leave for future research, but I think that with the data you have, you could already extract more detailed correlations. Moreover, sometimes I found it a little difficult to follow your text, because the figures are in the supplementary materials, and I think that your manuscript lacks a figure in which you show the GPR radargram, the photomosaic and the lidar depth together on the same transect.

I do not find severe problems in the manuscript, but it may be improved if the relation between instrument uncertainties and their causes is discussed in more detail. I suggest the manuscript to be accepted after minor revisions.

Specific comments:

Figure 1: Please, add a scalebar to panel f). Additionally, consider adding the location of Fairbanks in panel f), since it was mentioned several times in the text.

Data availability. Please, add a link to the NSIDC DAAC repository. Also, the fact that you put a part of data availability in the main text and a part in the supplementary material is confusing, in my opinion.

Chapter S1 Ground-Penetrating Radar Systems and Methods: consider adding more details about which instrument was used in which site. Especially, one of the instruments (the GSSI one) had worse GPS positioning than the others. It would be important to assess if this introduced greater uncertainties in the GPR-lidar comparison.

Figure S1: I think that you missed a great opportunity to show the DN013 radargram (and the lidar estimated depth) here. If you do so, you could elaborate on the buried vegatation that contributed to the GPR and lidar uncertainties (e.g. "at transect distance = 5 m, a buried little tree is also shown in the radargram as these hyperbola circled in red in the radargram, and this had this ... negative effect on the lidar snow depth estimation").

Figure S2: I am not convinced that the resolution difference you mention in the Figure caption is due to the different antenna frequency. In my experience, 1 GHz is already enough to show features as little as those recorded by the 1.6 GHz antenna, maybe just a little worse. I suppose that the perceived resolution difference between the two images is just due to the different spatial resolution: DB254 seems to have 1 trace per 10 cm, while SA326 has a much higher spatial resolution (I can't count the pixels, but they are much more than DB254). Also, similar to what I told for Figure S1, I really would like to see if the snowpack photomosaic of DB254 is different from that of SA326, to more constructively assess if the differences in the radargrams are related to the vegetation buried under the snowpack.

Figures S3 to S5: I would put at least one of them in the main text. Also, in the legend, please note on which date the lidar survey was performed, so it is easy to visually assess which transects were surveyed before and after the lidar.

Supplementary text, line 126: a full stop is missing at the end of the line.