Advances in mapping sub-canopy snow depth with unmanned aerial vehicles using structure from motion and lidar techniques

Vegetation has a tremendous influence on snow processes and snowpack dynamics yet remote sensing techniques to resolve the spatial variability of sub-canopy snow depth are lacking. Unmanned Aerial Vehicles (UAV) have had recent widespread application to capture high resolution information on snow processes and are herein applied to the sub-canopy 10 snow depth challenge. Previous demonstrations of snow depth mapping with UAV Structure from Motion (SfM) and airbornelidar have focussed on non-vegetated surfaces or reported large errors in the presence of vegetation. In contrast, UAV-lidar systems have high-density point clouds, measure returns from a wide range of scan angles, and so have a greater likelihood of successfully sensing the sub-canopy snow depth. The effectiveness of UAV-lidar and UAV-SfM in mapping snow depth in both open and forested terrain was tested in a 2019 field campaign in the Canadian Rockies Hydrological Observatory, Alberta 15 and at Canadian Prairie sites near Saskatoon, Saskatchewan, Canada. Only UAV-lidar could successfully measure the subcanopy snow surface with reliable sub-canopy point coverage, and consistent error metrics (RMSE <0.17m and bias -0.03m to -0.13m). Relative to UAV-lidar, UAV-SfM did not consistently sense the sub-canopy snow surface, the interpolation needed to account for point cloud gaps introduced interpolation artefacts, and error metrics demonstrate relatively large variability (RMSE <0.33m and bias 0.08 m to -0.14m). With the demonstration of sub-canopy snow depth mapping capabilities a number 20 of early applications are presented to showcase the ability of UAV-lidar to effectively quantify the many multiscale snow processes defining snowpack dynamics in mountain and prairie environments.

Line 66: Leading to variably, I think you mean variability Line 70: It would be great to reference  here. Table 1 in their paper reviews this and they provide their own evaluation metrics of ALS in a forest and open area. I would also reference . They showed a comparison of lidar in Switzerland to snow depth transects in forested areas as well.
Line 75: TLS was used in the forest in . Yes, the TLS did not go all the way into the entire forest but from an evaluation perspective of airbone lidar or SfM there's little difference from being 300 meters in a forest as long as there are consistent trees overhead that would inhibit returns from the laser. Also, their paper did not explicitly show that TLS couldn't be used further in the forest, it just gets more complicated.
Line 90: Could add that (Zheng et al., 2016) lidar to understand vegetation processes effect on snow. They particularly note bias that might occur due to tree wells. (Currier & Lundquist, 2018) used lidar to understand the snow-vegetation interactions in multiple climates.  also used airborne lidar data to improve the understanding of snow depth related to the forest in Colorado and Switzerland.
Line 190: I would mention here that the code is provided on your github page. Great job with providing this.
Line 205: Trees typically are taller than 50 cm. Most people consider a tree to be at least 2 m tall. Why did you choose 50 cm? This is inconsistent with what the caption shows in Figure 4.

Line 230: What is estimated and what is observed? I'd say UAV-derived Snow Depth and Snow
Depth Probe Manual Observations, or something more specific.
Line 235: Yes, the reported error metrics are inflated when moving into the forest. It'd be worthwhile mentioning that the sample size is much less. Some lidar points do great. In the methods the GNSS mentions a ±2.5 cm accuracy, how was that determined. Is it possible that this is inflated when in the forest? If not, mention that. Are these errors from how the point cloud was processed and points were classified? Is ±2.5 cm true for both horizontal and vertical accuracy?
Line 238: I'd start a new paragraph when introducing the error metrics with SfM.
Line 245: The authors should be using Digital Terrain Models instead of Digital Surface Models throughout.  Currier et al. 2019, that the airborne lidar is more likely to penetrate the shrubs than the TLS observations. What's the scientific name for the shrubs found at these locations?
Figures: I would change the easting northing to the total number of meters within the domain, or start at 0 and show ticks from 0 m. I don't know the projection information, and if I did the numbers aren't that meaningful. If the location is important, please provide the UTM zone. But still it's a bit annoying to do the subtraction each time to get a sense of scale. I would just make it easier for the readers, if possible. Otherwise the figures are great.
Line 317: This seems like an appropriate time to re-mention UAV lidars ability to capture tree wells.
Line 321: Confusing sentence. Deems reported errors in the forest larger than 14 cm? Why is 14 cm mentioned. Figure 5 reports RMSE of 0.15 and 0.16.
Also, in the previous sentence. Studies have masked out the forest? Studies have looked at airborne lidar accuracy in the forest.

Line 355: Really cool figure and analysis
Line 375: Green polygons look cyan when zoomed out, might choose a different color. Furthermore, the near infrared data seemingly comes out of nowhere -maybe provide some more context within the section for it and why it needs to be mentioned. Provide a citation for NIR serving as a proxy for albedo.
Line 435: "The accuracy and resolution demands mean that bare surface classification techniques suitable for airborne platforms that efficiently resolve topography and hydrography at watershed scales from last returns will be unsuitable for resolving the snow depth around a particular shrub from a dense point cloud for example" The paper did not show that using the last returns was unsuitable. The classification technique used something similar to last returns. Previous studies have showed using the last returns resulted in a generally unbiased snow depth estimate, and provided a reasonable approximation of the variability. I am not sure what this sentence is attempting to say.
Line 465: A discussion referencing the difficulties with modeling in Mark Raleigh's paper seems appropriate and a better citation then Tom Painters 2016 paper. Furthermore, when mentioning snow pack density variability, mentioning Karl Wetlaufer's paper seems appropriate (Raleigh & Small, 2017;Wetlaufer et al., 2016).
Line 479: "The UAV-lidar metrics consistently exceed the UAV-SfM metrics and are better than previously reported results in the airborne-lidar and UAV-SfM literature." This isn't true. Metrics are similar but not better than. Please note line 69.