Monitoring snow depth change across a range of landscapes with ephemeral snowpacks using structure from motion applied to lightweight unmanned aerial vehicle videos

Differencing of digital surface models derived from structure from motion (SfM) processing of airborne imagery has been used to produce snow depth (SD) maps with between similar to 2 and similar to 15 cm horizontal resolution and accuracies of +/- 10 cm over relatively flat surfaces with little or no vegetation and over alpine regions. This study builds on these findings by testing two hypotheses across a broader range of conditions: (i) that the vertical accuracy of SfM processing of imagery acquired by commercial low-cost unmanned aerial vehicle (UAV) systems can be adequately modelled using conventional photogrammetric theory and (ii) that SD change can be more accurately estimated by differencing snow-covered elevation surfaces rather than differencing a snow-covered and snow-free surface. A total of 71 UAV missions were flown over five sites, ranging from short grass to a regenerating forest, with ephemeral snowpacks. Point cloud geolocation performance agreed with photogrammetric theory that predicts uncertainty is proportional to UAV altitude and linearly related to horizontal uncertainty. The root-mean-square difference (RMSD) over the observation period, in comparison to the average of in situ measurements along similar to 50 m transects, ranged from 1.58 to 10.56 cm for weekly SD and from 2.54 to 8.68 cm for weekly SD change. RMSD was not related to microtopography as quantified by the snow-free surface roughness. SD change uncertainty was unrelated to vegetation cover but was dominated by outliers corresponding to rapid in situ melt or onset; the median absolute difference of SD change ranged from 0.65 to 2.71 cm. These results indicate that the accuracy of UAV-based estimates of weekly snow depth change was, excepting conditions with deep fresh snow, substantially better than for snow depth and was comparable to in situ methods.