Defining similar regions of snow in the Colorado River Basin using self-organizing maps

Abstract

We used self-organizing maps (SOMs) to define regions of homogeneity in the Colorado River Basin using snow telemetry (SNOTEL) snow water equivalent (SWE) data. SOMs are a specific application of artificial neural networks. Daily data for 216 stations using 15 years (1991-2005) of data from 1 October through 30 June were used. To identify areas of similar snow accumulation, persistence, and ablation patterns, data were transformed by dividing by the 15 year average peak SWE. Three experiments were performed to determine how the regions of homogeneous snowpack characteristics changed. The number of groups was increased from 4 to 6 to 9 to 16. By increasing this resolution, more subtle variations were defined. The temporal resolution of the data was decreased from daily to weekly to monthly to yearly. The accumulation and ablation of the snowpack over time represents a plot called a niveograph, which was summarized for yearly data by three variables. These were peak SWE, length of season, and date of peak SWE. Very similar results were produced using daily, weekly, and monthly time steps. However, using peak SWE produced only 50% of the same groupings, while using the other annual summary variables, even together, produced less than 25% of the same groupings. Using 18 physiographic variables to represent the SNOTEL stations yielded groups that were similar to those from using peak SWE but more evenly distributed in space. Using Ward's method of cluster analysis could only be performed with the individual annual summary variables. It produced groupings similar to the comparable SOM application but slightly less representative of the daily data groupings.