In situ monitoring and machine modeling of snowpack evolution in complex terrains

Abstract:

It is well known that snowpack evolution depends on variety of landscape conditions including tree cover, slope, wind exposure, etc. In this presentation we report on methods that combine modern in-situ sensor technologies with machine learning-based algorithms to obtain improved models of snowpack evolution. Snowcloud is an embedded data collection system for snow hydrology field research campaigns that leverages distributed wireless sensor network technology to provide data at low cost and high spatial-temporal resolution. The system is compact thus allowing it to be deployed readily within dense canopies and/or steep slopes. The system has demonstrated robustness for multiple-seasons of operation thus showing it is applicable to not only short-term strategic monitoring but extended studies as well. We have used data collected by Snowcloud deployments to develop improved models of snowpack evolution using genetic programming (GP). Such models can be used to augment existing sensor infrastructure to obtain better areal snow depth and snow-water equivalence estimations.

The presented work will discuss three multi-season deployments and present data (collected at 1-3 hour intervals and a multiple locations) on snowdepth variation throughout the season. The three deployment sites (Eastern Sierra Mountains, CA; Hubbard Brook Experimental Forest, NH; and Sulitjelma, Norway) are varied not only geographically but also terrain-wise within each small study area (~2.5 hectacre). We will also discuss models generated by inductive (GP) learning, including non-linear regression techniques and evaluation, and how short-term Snowcloud field campaigns can augment existing infrastructure.