Performance and Controlling Factors of Airborne LiDAR Snow Depth Estimates in Boreal Forests: Insights from NASA SnowEx 2023 Alaska Campaign

Liu, Jipeng; Cho, Eunsang; Vuyovich, Carrie M.

doi:10.5194/egusphere-2026-986

Preprints

https://doi.org/10.5194/egusphere-2026-986

Preprints

16 Jun 2026

| 16 Jun 2026

Performance and Controlling Factors of Airborne LiDAR Snow Depth Estimates in Boreal Forests: Insights from NASA SnowEx 2023 Alaska Campaign

Jipeng Liu, Eunsang Cho, and Carrie M. Vuyovich

Abstract. Quantifying spatial distribution of the snowpack is crucial for hydrological, ecological, and climate research, as well as their applications. Due to the high spatial resolution and extensive coverage, Airborne Light Detection and Ranging (LiDAR) has emerged as an effective tool for large-scale snow depth estimation. However, discrepancies between LiDAR-derived and manually measured snow depth values exist across areas influenced by topographical and vegetation characteristics such as canopy height, slope, and roughness. This study aims to 1) evaluate the performance of the airborne LiDAR snow depth measurements compared to magnaprobe in-situ data and 2) identify key factors affecting the accuracy of airborne LiDAR snow depth measurements focusing on the boreal forest environment. We utilize airborne LiDAR data and ground-based snow depth observations collected in the Fairbanks region of central Alaska during NASA SnowEx 2023 Alaska Campaign. The study focuses on three subregions: Bonanza Creek Experimental Forest (BCEF), Farmers Loop Creamers Field (FLCF), and Caribou-Poker Creeks Research Watershed (CPCRW). The results showed that the LiDAR snow depth data has a reasonable agreement with in-situ observations (R: 0.605, Mean Absolute Error: 18.8 cm) but exhibits varying levels of errors across the three subregions. By applying regression analysis and machine learning, we quantify the contribution of individual factors to measurement discrepancies and determine which factors are most influential. We employed Gradient Boosting Machine (GBM) model using five LiDAR-derived environmental variables—canopy height, elevation, slope, roughness, and ground point density—as predictors of relative error. Across all subregions and models, canopy height consistently emerged as the most important factor of LiDAR snow depth error.

Received: 20 Feb 2026 – Discussion started: 16 Jun 2026

Competing interests: At least one of the (co-)authors is a member of the editorial board of The Cryosphere.

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 2469 KB)

Supplement (565 KB)

Download & links

Jipeng Liu, Eunsang Cho, and Carrie M. Vuyovich

Status: final response (author comments only)

Subscribe to comment alert

RC1:
'Comment on egusphere-2026-986', Anonymous Referee #1, 25 Jun 2026
Referee comment of egusphere-2026-986 manuscript entitled "Performance and Controlling Factors of Airborne LiDAR Snow Depth Estimates in Boreal Forests: Insights from NASA SnowEx 2023 Alaska Campaign" by Liu et al.
General comments
In this manuscript, the authors present three different study sites in Alaska where NASA collected LiDAR snow depth data during SnowEx-campaigns in 2023. The aim is to examine a) how well these measurements correspond to reality in the area’s various land cover types, particularly in forests, and b) utilize machine learning to address which terrain features specifically affect the accurate measurement of snow depth. One goal is to provide quality control tools for remote sensing snow depth products. I believe this article has much to offer in this field of research, as similar studies specifically involving LiDAR data collected from an aircraft with flight height of 700m and field measurements on this scale at boreal sites are scarce. In this regard, systematic reporting of the data, methods and results would be essential. The research design is simple, which means time could be devoted to more detailed reporting.
At this stage, it is not entirely clear to the reader how the data has been processed, how the analysis was structured and conducted, and the results are not reported consistently or clearly. The figures could be combined, and I recommend creating a summary table to support the findings presented in the figures. Additionally, the manuscript contains claims for which I would like to see more references, such as the effect of ground frost on snow depth. In general, I find the references to be quite insufficient—yet there have been quite a few studies on this topic that could be cited and compared. In the "Discussion" section, I would definitely like to see more reflection on the analysis and its limitations, together with comparison to similar studies.
I believe that clarifying the text, clearly stating and describing the data and methods, adding more recent publications to the sources, and streamlining or combining the figures—as well as possibly presenting data in tables—would significantly improve the article’s readability and scientific value.
Specific comments
The abstract could do a better job of highlighting why this study is so important and what it actually found. After all, quite a few studies like this have already been conducted. Perhaps it could be marketed better: why is this study so important?

Magnaprobe is mentioned in the abstract, but never afterwards in the text. In the methods section, a manual probe is described. It should be correctly stated which instrument has been used.

When describing the previous LiDAR studies in the introduction, I would recommend clarifying the context at the beginning. LiDAR can easily achieve centimeter-level accuracy, and point density way higher than 1 point/m² depending on the vehicle and flight height and speed, overlap etc. I assume the authors are referring specifically to LiDAR mapping done from an airplane, not a drone. This could be clarified for the reader to make it clear that the study specifically refers to high-altitude ALS devices, not drones. Later in the introduction UAS are introduced, so I would restructure the chapter to clarify different devices and what accuracies there are in different ALS platforms.

In my opinion, the "Introduction" section could include additional sources that evaluate the accuracy of LiDAR in measuring snow depth, through different platforms (airplanes, helicopters, drones, terrestrial). Currently, there is an entire paragraph there that cites only NASA 2018 as its source—but that section discusses, for example, mapping alpine environments and remote sensing of snow in general.

There are no sources in the section describing the study sites, yet it describes weather and snowfall amounts. I would recommend including information on the years for which the averages were calculated, as well as the data source and location.

Section 3.1: In this section, it would be helpful to explain in more detail: where the point clouds were georeferenced, what was done with them afterward, what algorithm was used to calculate the DEMs, and in which software program. Later, there’s a reference to NASA. This could be clarified for readers who haven’t worked with this data before. Since metadata is available for Larsen 2024, one could, for example, note that additional information on the data processing can be found in this source.

Section 3.1: Areas where there were no valid return signals were removed or extrapolated only in March—was there no need to do this in the May data? Why?

The text makes extensive use of parentheses () to supplement sentence structures. While their use may be justified in some cases, I think there are far too many of them, which makes the text confusing. I recommend significantly reducing the use of parentheses and considering how the ideas could be expressed without them.

Figures 1 & 2 could be combined, now the figure 1 is a broad figure without too much information in practical, meanwhile figure 2 is giving a lot of important background information of the area itself. Figure 2 is not referred in any part of the main text.

Section 3.3: The sources are poorly cited. As I noted earlier in comment 5, in some sections they are not cited at all. But in this section, for example, Larsen 2024a and Larsen 2024b are not cited in the reference list in a way that makes it clear to the reader which source is being referred to. Overall, section 3.3 could be more concise without subheadings. References to Larsen could be cited more clearly. This section discusses relevant points regarding why these parameters were chosen, but these could perhaps be addressed elsewhere rather than in this "Data and Methods" section. Variables in equation 2&3 are not explained.

Section 3.5: To support the choice of this method, we would need references explaining why it is so well-suited for a dataset of this size (300 points per site?). Additional information about the method, such as training data, testing, and validation sample sizes, would be important.

Figures 4-6: It is not clear to the reader what these four sub-areas actually are. The captions are repetitively unclear; it consists of four lines that repeat the same information twice. I’d cut the number of these maps in half—there are currently three, and each one has four areas that don’t really provide any context for the maps. It would work just fine to have, say, two examples per page.

Section 4.2: It might be a good idea to create a summary table of the results, so you can see, for example, the sample sizes across the different groups. Right now, it's not very easy for the reader to understand the differences between the groups. Once again, parentheses are used, which clutter up the text and bring up points that have already been explained. The phrase "markedly higher proportions" is used, even though the differences are not necessarily that large; there is also variation within them, which may be related to the method itself. For example, the largest proportion of very high relative error is in the 10–15 m class, but the error decreases in the other, larger tree size classes. Text is also referring to figure 8, but figure 8 has also a, b and c sub-figures. The figure captions are showing information that should be explained in the text rather than in the captions. Overall, when looking at this analysis, one might wonder why elevation was even included as an explanatory factor in the first place. The reasons behind these choices need to be explained more clearly. The section also claims that “Higher density areas tended to have smaller errors, suggesting that sufficient ground returns improve terrain characterization beneath the snowpack”—of course, more points produce a more accurate terrain model. And in that case, there’s no need to even see beneath the snow. It sees the surface it’s mapping. The statement doesn’t really make sense.

Section 4.3: It says here that this is a reasonable result. But what is that result it is referring to? The CPCRW region is also significantly larger in area than the others, which might also affect the analysis?

Figure 11: In this figure, the captions don't explain differences between a/b/c/d. It is not clear, what are the two lines labeled 2 and 10 in c-plot, or what are those lines at the bottom of each subplot? For example, in that c-plot, they look completely different from the ones in the others. It would probably be enough if that labels were just on the X-axis and in the figure caption – currently each subplot has a title.

Section 5: In the Discussion section, I would like to see references to other studies conducted using aircraft or other SnowEx-campaigns. Currently, the accuracy measurements are being compared to the study by Jacobs et al., which was conducted using a drone. Once again, there are paragraphs without a single source. Section 5.2 begins with the claim: “Relatively few studies have directly quantified airborne LiDAR snow depth accuracy in dense forests,” which is not necessarily entirely true, as studies have indeed been conducted—admittedly not with a specific focus on densely forested areas—but there is nevertheless evidence that accuracy is reduced due to the influence of vegetation. In my opinion, this section lacks essential studies conducted after the 2020s. In the discussion section, I think the role of terrain variables in the analysis, as well as the chosen modeling method and its limitations, could be examined in greater detail. For example, regarding slope, there is no mention at all of the fact that slope direction is a proven significant factor in snow distribution. Furthermore, the small sample size is not addressed, nor is GBM, for that matter.

Section 5.3: Here is where I would expect the issues of the comment 16 to be discussed, but it focuses on addressing the accuracy of field measurements. Of course, there is uncertainty involved, but take, for example, the statement: "For example, under a tall tree, snow depth can vary within a few meters due to drip zones, microtopography, or vegetation cover. A single probe might read 40 cm, but just a few meters away the depth could be 30 cm.”— it is true, but that is precisely why it is important to take many manual measurements, locate them precisely, and exactly why a combination of field + remote sensing data collection is important. Furthermore, at this point in the discussion, new information is presented regarding GPS accuracy, which should have been highlighted earlier in the “Data and Methods” section: I find it strange that GPS accuracy varies between 2.5 m and 15 m. Manual measurements should be very precise to serve as “ground truth,” but here it seems that LiDAR-based data with 1-meter resolution is significantly more accurate than manual measurements with a range of 2–15 meters horizontal accuracy.

Technical corrections (non-exhaustive)
Line 39-40: I would clarify this sentence, for example, how do the terrain and vegetation affect the snow accumulation and therefore streamflow timing?

Line 49-51: What do you mean by “ground-based sensors cannot see the snow under the trees”?

Line 52: Sources?

Line 72-87: In this section, things get a bit jumbled: vegetation is mentioned here and there, and the discussion of topography, landforms, sublimation, and tree cover is a bit muddled. It could be clarified, and the sources cited more clearly. Additionally, the section begins by discussing how the accuracy of the LiDAR depends on vegetation, but in the same section, it also discusses where snow itself accumulates. Perhaps these topics could be separated more clearly.

Line 91: What does it mean, when discontinuous permafrost affects snow accumulation and LiDAR performance? These sentences could be clarified, and they would need decent sources.

Line 93: Since not all readers may know what the NASA SnowEX 2023 campaign is, I think it would be helpful to describe it here, at least briefly.

Line 111: Figure 1 --> Fig. 1?

Line 116: What are those little red dots in the maps? What is the m? Elevation (from sea level?)? I'd also recommend removing the repetition from the image and the caption—right now the names are written in both.

Line 125: It might be a good idea to mention the size of the other regions as well

Line 131: Creamer’s field and Farmers Loop abbreviations are opened here, even though the FLCF has been used before. It would be reasonable to open the meaning of abbreviations once in the beginning and use them for the rest of the text.

Line 135: Maybe a new section for the climatic conditions?

Line 150: It might be a good idea to include the manufacturer and country of origin.

Line 160: dates for May 2022?

Line 163: Repetition

Line 165: I feel like “Thanks to” is now very scientific language – consider changing the phrasing?

Line 200: The equation could be written so that it is first mentioned in the text and then as a separate line of text, so it is easy to refer to it later.

Line 217: It's enough to explain these abbreviations just once, at the beginning. Here it’s "DTM" later it is again explained in Line 226.

Line 296: Figure 3 --> Fig. 3?

Line 311: FLCF abbreviation is opened again. It has been explained before.

Line 313: LiDAR abbreviation is opened again. It has been explained before.

Lines 310-313: Same sentence twice.

Line 334: BCEF abbreviation is opened again. It has been explained before.

Line 336: LiDAR abbreviation is opened again. It has been explained before.

Lines 333-336: Same sentence twice.

Line 342: CPCRW abbreviation is opened again. It has been explained before.

Line 344: LiDAR abbreviation is opened again. It has been explained before.

Lines 341-344: Same sentence twice.

Line 349: Refer to the equation.

Line 361: There is an underscore in the image legend

Line 407: Too many repetitive abbreviations in the caption. 9c does not have x-axis label.

Lines 442-448: In my opinion, this section contains information about the colors on the maps that would be suitable for a caption, whereas the caption (Line 455) presents the results of the analysis.

Line 601: Reconsider rephrasing: “meaning we can predict where LiDAR will have trouble”
Citation: https://doi.org/10.5194/egusphere-2026-986-RC1
RC2: 'Comment on egusphere-2026-986', Anonymous Referee #2, 21 Jul 2026

Liu et al. investigate lidar snow depth error from lidar and manual samples collected during the NASA SnowEx 2023 campaign. The authors investigate the sources of lidar error using canopy height, surface roughness, point density, elevation, and slope. They also generate a machine learning framework to estimate expected error using these variables. This work is an important contribution to the field given the emergence of lidar used for validation, while lidar itself is difficult to validate.
Prior to publication, the presentation of the results should be improved and several conclusions should be softened. The manuscript also currently lacks any discussion of in-situ sampling bias, which may be influencing the reported results. I believe this manuscript will be a valuable contribution to the field, but it requires significant revision before publication.
General comments:
A critical piece missing from this manuscript is a discussion of potential bias in the in-situ measurements used as ground truth. The manuscript assumes that manual snow probe measurements represent unbiased ground truth, yet the cited literature using the same dataset demonstrates that this assumption may not hold. Stuefer et al. (2025) demonstrate a positive bias of approximately 10 cm using the same SnowEx dataset (Figure 8), and Figure 9 of that study shows vegetation-induced air pockets that may explain this bias. Given that this manuscript reports a systematic negative bias in the lidar estimates, these previous findings should be discussed in much greater detail. At minimum, the authors should evaluate how this known sampling bias affects their conclusions. I would also encourage the authors to repeat their analyses after correcting for the reported 10 cm bias to determine whether the inferred lidar errors and variable importance change.
My second major concern is that several of the conclusions are stronger than the presented evidence supports. The manuscript repeatedly concludes that canopy height is the dominant control on lidar snow depth error and that steep slopes also increase errors. While canopy height does appear to be the strongest predictor among the variables considered, many of the figures do not convincingly demonstrate these relationships. For example, Figures 4-6 do not clearly show that larger errors occur beneath dense canopy, and the evidence for increasing error with slope is inconsistent across sites. Likewise, the relatively modest model skill indicates that a large fraction of the error remains unexplained. The discussion should better acknowledge that these variables explain only part of the observed error and that other factors, such as in-situ measurement uncertainty, snow-off lidar uncertainty, vegetation beneath the snowpack, or geolocation error, may contribute substantially.
The machine learning results would also benefit from additional discussion. If the predictive error of the GBM model is similar to the observed lidar error itself, then the practical utility of the model is unclear. Rather than emphasizing the predictive capability of the model, I think the more important conclusion is that lidar errors are difficult to predict using the available environmental variables. The manuscript briefly acknowledges the existence of unmodeled factors, but this deserves a much more thorough discussion. I also suggest testing snow depth as an additional predictor. Whether predicting absolute or relative error, snow depth itself may explain some of the scaling in the error and could improve model performance.
Finally, I think the presentation of the results could be strengthened. The manuscript currently focuses on maps of the study areas, but these figures are not especially effective at supporting the conclusions. Several additional summary figures would make the results much easier to interpret. For example, site-specific error statistics, probability density functions comparing lidar and in-situ snow depth, and distributions of canopy height or other predictor variables would provide a clearer picture of where the errors originate and whether the lidar observations reproduce the observed snow depth distributions. Throughout the manuscript, the systematic negative bias appears to be one of the primary results, yet bias receives much less emphasis than correlation or MAD.
Line by line comments
Title: consider adding in-situ / manual sampling to the title since this is the other key dataset aside from lidar.
19: Change ‘showed’ to present tense
41-51: Starting a paragraph with ‘however’ feels a bit awkward. Also, the focus of this paragraph could be improved. The first half is about snow sampling methods and the second have is about difficulty of sampling in forests. Maybe split into two paragraphs or improve the topic sentence / transitions.
56-57: Does this need a citation?
62-63: ‘ability to capture snow depth spatial variability that is unattainable with traditional methods’ is vague.
82: Awkward transition between sentences. Introduce that you are now talking about vegetation effects.
86-87: ‘even small absolute errors can be a large relative fraction’. I would remove this. This is an error metric / statistics phenomenon, not an issue with snow sampling.
Figure 1: label the color bars as elevation
Figure 2: Could you add a satellite image of the sites here or in figure 1? Would be helpful to conceptualize this sites more.
250: Suggest adding bias as a metric.
269-270: Did you attempt the ML framework with absolute errors instead of relative? I think this could potentially perform better, especially if providing snow depth as a predictor.
273: ‘blowing up the fraction’ seems informal. How about ‘disproportionately amplifying relative error magnitudes’
Figure 3: Consider reporting statistics separately for each site, either by splitting the figure into three panels or by adding a table. Bias should also be included as a summary statistic. The systematic negative bias appears to be one of the primary findings.
326-327: Figure 6 doesn’t explicitly support this claim
Figure 4-6. These figures provide useful context but do not clearly support the conclusions regarding canopy effects. Consider combining them into a single summary figure and replacing the remaining space with quantitative analyses, such as canopy height versus error.
Figure 7: I would also be interested in seeing this analysis repeated using absolute error rather than relative error. Some of the apparent increase in relative error under taller canopy may simply reflect shallower snowpacks.
364-367: I’m not convinced that the results demonstrate that errors increase with slope. Yes, there is a maximum error for FLCF, but not the two other sites.
Figure 9: label the x-axis for point density
426-427: ‘given the inherent variability and some unmodeled factors… this is a reasonable performance.’ This sentence is vague. Also, stating that it’s a reasonable performance is subjective.
449: ‘the modeling results reenforce that canopy height is the dominant control on lidar snow depth error’. I would specify that canopy is the dominant control out of the tested variables. The R² values for the model demonstrate that much of the error is not accounted for by the variables.
465-466: Providing the MAD of 19 cm for all points does not demonstrate that performance decreases under forest canopies. This statement would require two different errors for open and forested errors showing that.

Citation: https://doi.org/10.5194/egusphere-2026-986-RC2

Jipeng Liu, Eunsang Cho, and Carrie M. Vuyovich

Supplement

https://doi.org/10.5194/egusphere-2026-986-supplement

Jipeng Liu, Eunsang Cho, and Carrie M. Vuyovich

Viewed

Total article views: 71 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
35	27	9	71	8	6	4

HTML: 35
PDF: 27
XML: 9
Total: 71
Supplement: 8
BibTeX: 6
EndNote: 4

Views and downloads (calculated since 16 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	35	27	9	71

Cumulative views and downloads (calculated since 16 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	35	27	9	71

Viewed (geographical distribution)

Total article views: 54 (including HTML, PDF, and XML) Thereof 54 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 29 Jul 2026

Short summary

We analyzed airborne LiDAR data from NASA’s SnowEx 2023 campaign to study snow depth in Alaska’s boreal forests. Comparing LiDAR to ground measurements, we found it captures general snow patterns well, though tall trees and slopes reduce accuracy. This approach can help improve snow surveys and better manage water resources in remote forested areas.


Total:	0
HTML:	0
PDF:	0
XML:	0