UC Berkeley

The Sierra Nevada winter snowpack is the major water resource for the state of California. To better quantify the input of the water system, we deployed wireless-sensor networks across several basins in the Sierra Nevada. Together with operational and scientific research agencies, we also collected numerous scans of snow-on and snow-off lidar data over several basins in the high Sierra. We mined the lidar data and found how spatial patterns of snow depth are affected by topography and vegetation while elevation is the primary variable, other lidar-derived attributes slope, aspect, northness, canopy penetration fraction explained much of the remaining variance. By segmenting the vegetation into individual trees using lidar point clouds, we were able to extract tree wells from the high resolution snow-depth maps and we found the spatial snow distribution to be affected by the interactions of terrain and canopies. The snowpack is deeper at the downslope direction from the tree bole, however the snowpack at upslope direction being deeper. On sub-meter to meter scales, non-parametric machine-learning models, such as the extra-gradient boosting and the random-forest model, were found to be effective in predicting snow depth in both open and under-canopy areas. At spatial scales that are larger than 100 × 100 m2, we developed a novel approach of using the k-NN algorithm to combine the real-time wireless-sensor-network data with historical spatial products to estimate snow water equivalent spatially. The results suggest only a few historical snow-water-equivalent maps are needed if the historical maps can accurately represent the spatial distribution of snow water equivalent. The residual from the k-NN estimates can be distributed spatially using a Gaussian-process regression model. The entire estimation process can explain 90% of the variability of the spatial SWE.