Improving Snow Water Equivalent Maps With Machine Learning of Snow Survey and Lidar Measurements

Abstract

In the semiarid interior western USA, where a majority of surface water supply comes from mountain forests, high‐resolution aerial lidar‐based surveys are commonly used to study snow. These surveys provide rich information about snow depth, but they are usually not accompanied with spatially explicit measurements of snow density, which leads to uncertainty in the estimation of snow water equivalent (SWE). In this study, we use a novel approach to distribute ~300 field measurements of snow density with artificial neural networks. We combine the resulting density maps with aerial lidar snow depth measurements, bias corrected with a very large and precisely geolocated array of field‐measured snow depths (~4,000 observations), to create and validate maps of snow depth, snow density, and SWE over two sites along Arizona's Mogollon Rim in February and March 2017. These maps show differences between midwinter and late‐winter snow conditions. In particular, compared to that of snow depth, the spatial variability of snow density is smaller for the later snow survey than the earlier snow survey. These gridded data also show that the representativeness of Snow Telemetry and other point measurements is different for the midwinter and late‐winter snow surveys. Overall, the lidar artificial neural network SWE estimates can be as much as 30% different than if Snow Telemetry density were used with lidar snow depths to estimate SWE.

Forest cover and topography regulate the thin, ephemeral snowpacks of the semiarid Southwest United States

“…The second (“Salt”) box, located ~50 km east of Show Low, AZ, in the Salt River watershed, has higher elevations (~2,400–2,950 m a.s.l. ), generally less forest cover (due to the presence of montane meadows as well as a significant amount of burned area from a 2011 wildfire; Kennedy & Johnson, ), and has a colder and drier climate than the Verde box (Broxton, van Leeuwen, & Biederman, ).…”

“…These SWE and snow depth maps span a large range of forest structures and topographic conditions for each area on three dates, before peak SWE in 2017 (February 1, 2017), near peak SWE in 2017 (March 7, 2017), and near peak SWE in 2019 (March 4, 2019). Broxton, van Leeuwen, & Biederman () describe the creation of these maps using the 2017 lidar and field data for the Salt and Verde boxes. The same methodologies are applied to the 2019 data.…”

“…The SWE maps, based on the lidar snow depth data and field snow density measurements (Section 2.2), are rich data sets for assessing the spatial variability of snowpack, but it is difficult to isolate the snow variability that is caused by forest cover, because so much of the snow variability is related to topography. For these study sites, elevation and northness (sin [slope] × sin [aspect]) are the leading physiographic indicators of SWE (Broxton, van Leeuwen, & Biederman, ). Vegetation influences, which are very important drivers of SWE variability at the local scale (ones to hundreds of metres), are second order to these terrain influences at larger scales (ones to tens of kilometres; Broxton et al, ) making it hard to isolate them across the lidar domains.…”

“…For these study sites, elevation and northness (sin [slope] × sin [aspect]) are the leading physiographic indicators of SWE (Broxton, van Leeuwen, & Biederman, ). Vegetation influences, which are very important drivers of SWE variability at the local scale (ones to hundreds of metres), are second order to these terrain influences at larger scales (ones to tens of kilometres; Broxton et al, ) making it hard to isolate them across the lidar domains.…”

Shallow snow depth mapping with unmanned aerial systems lidar observations: A case study in Durham, New Hampshire, United States

“…In contrast, ALS surveys typically report surface model densities between 8 to 16 points/m 2 (Broxton et al, 2015;Currier et al, 2019;Kirchner et al, 2014) and ground returns between 3 and 6 points/m 2 (Broxton et al, 2019;Kirchner et 340 al., 2014). ALS derived snow depth maps have a much greater proportion of areas that are masked due to no ground returns, particularly under trees, with masking areas ranging from less to 10% to more than 23% (Harpold et al, 2014;.…”