The manuscript “Spatially distributed snow depth, bulk density, and snow water equivalent from ground-based and airborne sensor integration at Grand Mesa, Colorado, USA”, uses ground penetrating radar (GPR) and airborne lidar to infer snow density during the February 2020 NASA snow-ex campaign.  The authors implement an automated GPR data processing workflow, and then evaluate them using field data.  They then go on to generate continuous snow density maps using multilinear regression and a couple of machine learning approaches, calculate the length scale of SWE, depth, and density, and attribute the variability of snow density to physical processes (in particular, highlighting the role of wind effects).  Overall, this manuscript is well written with detailed analyses, though I have some concerns about lack of focus and bringing out the novel aspects of the study, as well as quite a few other, mostly minor comments. 

Major Comments:

Overall, the setup is a little bit muddled.  The introduction and abstract both start out talking about spaceborne remote sensing, though the link between the snow density measurements at the field site and space-based remote sensing is unclear.  Is it about providing validation data, information about variability of snow density and SWE, radar technology / methodologies that can someday be useful for satellite missions, or what?

While the manuscript is very detailed and the quality of the analyses seems high, it is a little unfocused.  There are a lot of really interesting analyses (linking the GPR/Lidar measurements - including the ability to quickly analyze the data, the regression modelling of snow density, linking the maps to wind, terrain, and vegetation effects, and looking at things like the length scales of snow density and SWE and evaluating the uncertainty of the estimates).  However, it is a little hard to know what, among these analyses is the most important / novel.  For example, the regression modelling approaches have been applied to field snow density measurements elsewhere (though not with GPR) - and length scale evaluation is not new (though interesting because you did it for SWE, depth, and density).  These results, as well as attribution of snow density fields to wind patterns and vegetation, also seem pretty site specific.

The discussion section is also quite long and wanders a lot.  It mirrors the results in that it jumps from topic to topic.  It would probably be more effective to have more linkages between the topics and put them in a broader context - and keep linking back to a set of key points or research questions.

Overall, I would suggest reducing the manuscript - at least the parts that aren’t the main focus - and making clear what is novel vs what is auxiliary.  It would also help to think about a clear problem statement, research questions, hypotheses, etc and keeping the manuscript focused around those things.  Without this focus, it is hard to judge its novelty.

Specific Comments

Line 24: How does this study represent the “next step towards broad-scale SWE retrieval”.  How does broad-scale SWE retrieval depend on the methods / results shown here?

Introduction: There is dissonance between the initial setup of the study (which is about working toward the improvement of space-based remote sensing and talking about snowpack measurement and mapping at the local scale (e.g. which is measured by GPR, airborne lidar, drones, and field measurements).

Line 34: I think this comment should be revised to “simple spatial interpolation from ground observations is difficult” - things like physiographic variables and past snowpack patterns can be useful for distributing snow measurements.

Line 39: While many of these technologies have, in some circumstances produced fairly accurate maps, there are many caveats and there are also many cases where they fail, or are severely biased

Line 42: I do not understand what is being said here.  The first and second parts of this sentence do not seem to be related.

Line 81: This paragraph is long and disorganized.  The topic sentence is about the spatial patterns of snow density being related to underlying processes, followed by a listing of (sometimes and sometimes not) spaceborne techniques that could be useful for snow density measurement .

Line 89: If singling out ground temperature and roughness length, please explain why they are important in this application

Line 97: Drone based photogrammetry to get snow depth?

What is the relationship between GPR measurements to get snow density and space based remote sensing?  Can this technique be used to detect snow density using satellite data?

Line 104: Rather than simply giving a high level overview of your methods, it would be nice to frame this more in terms of research questions, hypotheses, and knowledge gaps.

Line 140: Remove the word “whereas”

Line 142-143: “the many forested stands in the survey domain” - this is pretty wordy.  Suggest just saying “forested stands”\

Why wasn’t the same type of instrument used in open and forested areas?  How might this have affected results? 

Line 147: Was the same type GPS used on both GPR units (with lower accuracy in the forest)?

Line 160: Were these processing steps the same for both GPR units

Line 213: “Time series” - this is pretty vague…and perhaps not even necessary since you are only using the Feb 1^st flight

Line 215: What is the principle of the Štroner et al method?  It is probably better, in the interest of having the work stand on its own, to be somewhat descriptive of all the methods, rather than just referring readers to other papers.

Line 230: Change “the” to “in”

Line 250-254 - This passage is confusing.  How were you able to determine the new random error?  It also seems pretty generous to call half the sample (outside the 25^th-75^th percentiles) outliers.

Line 270 - Can you be a little more descriptive of how the threshold was determined.

Line 293-301: This passage is confusing.  Can it be rephrased / reduced?

Line 340: Does this mean that the only snow depth differences correctly captured by the lidar was the difference between canopy and open areas.  Could this be commented on further in the discussion section (e.g. could it be something related to geolocation errors, differences in the scale of measurements, relatively large lidar errors relative to the variability in either of those environments individually, etc)?

Line 350: Maybe list in the text what this correlation is

Line 351-353: “The standard deviation …” - this sentence is awkwardly phrased

Line 354: Remove “uniquely”

Line 368-369: “the relative length-scales of variability for SWE closely resemble that of TWT and indicate that TWT is a better informer of SWE than either depth or density, independently” - this is confusing.  Is there a way to clarify why a smaller difference in length-scale makes TWT a better informer of SWE

Line 378: “Average of regression models” - was there just one ANN, RF, and ML model, so that the average is just the average of 3 models, or were there, for instance, an ensemble of ANN models that were averaged together?

Line 405: To get SWE here, do you use average snow density from all pits, as well as the random field field generated from all field density data?  It doesn’t seem like the SWE comparisons are fair if some methodologies have access to the validation data (e.g. average snow density) and some do not (e.g. regression maps, as far as I can tell).

Line 408-409: “depth is primary to SWE in this environment” - what does this mean?

Line 409-410: “Assessed at the average value of SWE for all 96 snow pits” - this is also unclear

Line 447: 2% - are the units of this supposed to be comparable to the previous number (1 kg/m3)?

Line 449-450: It is hard to reconcile how the errors in SWE found using this methodology at this field site relate to a global SWE uncertainty target.

Line 539: Could you please define the acronym “FMCW”?

Line 555: Remove comma after SWE.  Also, I find this statement confusing.

Line 562: Though by comparing SWE estimates at independent snow pits, it seems like you can surmise which method probably performs better.

Line 567: Which retrieval methodologies are you referring to here?

Line 590: I don’t think there is a regression learner toolbox.  The Regression Learner app is part of the statistics and machine learning toolbox…though neural networks are not part of these tools.   How did you implement the neural network models?

See pdf.

In this manuscript, the authors present a method to invert spatially distributed estimates of bulk snow density from the combination of ground-penetrating radar (GPR) retrievals and snow depths generated from airborne lidar. The bulk density map is then combined with a snow depth map at 1 m spatial resolution to calculate spatially distributed SWE for February 1, 2020. The snow density, snow depth and SWE maps are evaluated with field measurements from the NASA SnowEx campaign at Grand Mesa (Colorado, USA) during February 2020. These measurements span forested, exposed, and forest-adjacent areas over relatively flat subalpine terrain during cold and dry winter conditions.

The novel component of this manuscript appears to lie primarily with the generation of spatially distributed bulk snow density estimates from the GPR data using airborne snow depth maps to make the inversion. While algorithms and techniques currently exist for the retrieval of bulk snow density from GPR data (Griessinger et al., 2018; McGrath et al., 2022) the authors present new techniques for automated layer picking and introduce the use of snow depths from airborne lidar. The resulting distributed snow density estimates (after filtering of outliers) show similar accuracies to previous studies (Griessinger et al., 2018) when compared to high-quality snow density measurements, with an RMSE of 27 kg/m³.

The spatial patterns of bulk snow density presented are certainly revealing and will be of great interest to the snow research community. The implications of this variability on SWE retrievals speaks directly to the NASA SnowEx project goals, is an advancement in our understanding, and will be useful for the broader SnowEx community. However, the study misses an opportunity to compare the newly generated SWE maps to existing SWE maps such as those from the Airborne Snow Observatory(ies), which were produced for 1-2 Feb 2020 over the Grand Mesa – using modeled bulk snow density. With this, I think the study falls short of demonstrating improved SWE mapping (using GPR bulk density observations) by failing to compare (and contrast) the newly produced SWE product with existing data ASO that currently represents best practices in SWE mapping (despite the larger spatial resolution of 50m), and potentially other SWE estimates generated as a part of the SnowEx effort. Despite the main thrust of the paper being the GPR bulk density measurements and methods, I propose that the true scientific advancement of this work lies with such a SWE comparison.

Finally, while I appreciate the breadth of observations and careful use of many sources of field measurements that were used in the study, I found the manuscript narrative to be unfocused at times. Perhaps there is opportunity to reduce scope to include only the pertinent analyses and discussion that support the main purpose of the manuscript.

See attachment for further detail and line comments.