Terrestrial and airborne laser scanning and structure from motion techniques have emerged as viable methods to map snow depths. While these systems have advanced snow hydrology, these techniques have noted limitations in either horizontal or vertical resolution. Lidar on an unpiloted aerial vehicle (UAV) is another potential method to observe field- and slope-scale variations at the vertical resolutions needed to resolve local variations in snowpack depth and to quantify snow depth when snowpacks are shallow. This paper provides some of the earliest snow depth mapping results on the landscape scale that were measured using lidar on a UAV. The system, which uses modest-cost, commercially available components, was assessed in a mixed deciduous and coniferous forest and open field for a thin snowpack (< 20 cm). The lidar-classified point clouds had an average of 90 and 364 points/m² ground returns in the forest and field, respectively. In the field, in situ and lidar mean snow depths, at 0.4 m horizontal resolution, had a mean absolute difference of 0.96 cm and a root mean square error of 1.22 cm. At 1 m horizontal resolution, the field snow depth confidence intervals were consistently less than 1 cm. The forest areas had reduced performance with a mean absolute difference of 9.6 cm, a root mean square error of 10.5 cm, and an average one-sided confidence interval of 3.5 cm. Although the mean lidar snow depths were only 10.3 cm in the field and 6.0 cm in the forest, a pairwise Steel–Dwass test showed that snow depths were significantly different between the coniferous forest, the deciduous forest, and the field land covers (p < 0.0001). Snow depths were shallower, and snow depth confidence intervals were higher in areas with steep slopes. Results of this study suggest that performance depends on both the point cloud density, which can be increased or decreased by modifying the flight plan over different vegetation types, and the grid cell variability that depends on site surface conditions.

Observation‐based long‐term gridded snow water equivalent (SWE) products are important assets for hydrological and climate research. However, an evaluation of the currently available SWE products has been limited due to the lack of independent SWE data that extend over a large range of environmental conditions. In this study, three daily long‐term SWE products (Special Sensor Microwave Imager and Sounder [SSMI/S] SWE, GlobSnow‐2 SWE, and University of Arizona [UA] SWE) were evaluated by seasonal snow cover and land cover classifications over the conterminous United States from 1982 to 2017, using the historical airborne gamma radiation SWE observations (20,738 measurements). We found that there are similar patterns in SSMI/S and GlobSnow‐2 SWE when compared against the gamma SWE. However, GlobSnow‐2 SWE had better agreement with gamma SWE than SSMI/S SWE in some forested‐type classes and maritime and prairie snow classes. As compared to SSMI/S and GlobSnow‐2 SWE, UA SWE has much better agreement with gamma SWE in all land cover types and snow classes. Tree cover and topographic heterogeneity affect the agreement between the gamma and gridded SWE and accuracy of gamma SWE itself with the largest differences typically occurring when the percent tree cover was 80% or higher, the terrain slope was steeper than 2.5°, and the elevation range exceeded 100 m. The results demonstrate the reliability of the UA SWE products and the benefits of the gamma radiation approach to measure SWE, especially in forested regions.

Seasonal snow is an essential component of regional and global water and energy cycles, particularly in snow-dominant regions that rely on snowmelt for water resources. Land surface models (LSMs) are a common approach for developing spatially and temporally complete estimates of snow water equivalent (SWE) and hydrologic variables at a large scale. However, the accuracy of the LSM-based SWE outputs is limited and unclear by mixed factors such as uncertainties in the meteorological boundary conditions and the model physics. In this study, we assess the SWE, snowfall, precipitation, and air temperature products from a 12-member ensemble – with four LSMs and three meteorological forcings – using automated SWE, precipitation, and temperature observations from 809 Snowpack Telemetry stations over the western US. Results show that the mean annual maximum LSM SWE is underestimated by 268 mm. The timing of peak SWE from the LSMs is on average 36 d earlier than that of the observations. By the date of peak SWE, winter accumulated precipitation is underestimated (forcings mean: 485 mm vs. stations: 690 mm). In addition, the precipitation partitioning physics generates different snowfall estimates by an average of 113 mm with the same forcing data. Even though there are widespread cold biases (up to 3 ^∘C) in the temperature forcings, larger ablations and lower ratios of SWE to total precipitation are found even in the accumulation period, indicating that melting physics in LSMs drives some SWE uncertainties. Based on the principal component analysis, we find that precipitation bias and partitioning methods have a large contribution to the first principal component, which accounts for about half of the total variance. The results provide insights into prioritizing strategies to improve SWE estimates from LSMs for hydrologic applications.

Human‐induced landscape changes affect hydrologic responses (e.g., floods) that can be detected from a suite of satellite and model data sets. Tapping these vast data sets using machine learning algorithms can produce critically important and accurate insights. In the Red River of the North Basin in the United States, agricultural subsurface drainage (SD; so‐called tile drainage) systems have greatly increased since the late 1990s. Over this period, river flow in the Red River has markedly increased and 6 of 13 major floods during the past century have occurred in the past two decades. The impact of SD systems on river flow is elusive because there are surprisingly few SD records in the United States. In this study, Random Forest machine learning (RFML) classification method running on Google Earth Engine's cloud computing platform was able to capture SD within a field (30 m) and its expansion over time for a large watershed (>100,000 km²). The resulting RFML classifier drew from operational multiple satellites and model data sets (total 14 variables with 36 layers including vegetation, land cover, soil properties, and climate variables). The classifier identified soil properties and land surface temperature to be the strongest predictors of SD. The maps agreed well with SD permit records (overall accuracies of 76.9–87.0%) and corresponded with subwatershed‐level statistics (r = 0.77–0.96). It is expected that the maps produced with this data‐intensive machine learning approach will help water resource managers to assess the hydrological impact from SD expansion and improve flood predictions in SD‐dominated regions.