A New Study of Mountain Snowpack through Graph Spectral Analysis

The ability to accurately quantify snowpack variability has major impacts on estimations of snow depth and snow water equivalent (SWE). Complex physics based models have attempted to constrain the physical mechanisms that most affect snowpack properties and have been successful in producing results that are comparable to the average changes seasonal snowpack undergoes during the accumulation and melt periods. However they fail to accurately model any abrupt changes, especially in mountainous terrain; in part, due to the lack of in-situ measurements that contain the detailed forcing data required by such models. This study approaches the problem using a merger of Graph Theory, Spectral and K-means Cluster Analysis, as well as Radial Basis Function Neural Network to partition a complete data set of SNODAS product in addition to LiDAR datasets at the 1 km scale. These partitions are used to implement machine learning techniques to make predictions about current and future snowpack conditions. This system leverages the ability of Graphs, collections of nodes and edges, to represent nonlinear systems, such as the continuous spatiotemporal changes in seasonal snowpack and has unique properties that allow optimal sampling design for monitoring and studying snowpack patterns.