Characterizing and Predicting Watershed-Wide Snowpack Ripening Patterns with Machine Learning Methods

A melting snowpack initiates runoff production after cold content has been eliminated and the pore liquid water content has grown to overcome capillary resistance, a process called ripening. Here, we quantify the time-space distribution of ripening within a 4341 km² mountain basin in Montana, USA. Using model output for a 19 year period we compute a time-series of the energy needed for ripening, termed the Runoff Energy Hurdle (REH). The REH is associated with snowpack mass but is variably influenced by cold content, peaks earlier than mass, and is typically eliminated in days. We show that individual locations have complex year-to-year histories of REH growth and loss. Through K-means clustering, we identify four distinct ripening behaviors across high year-to-year variability. One cluster has ripening events throughout the snow season and can include 7-92 % of the basin depending on the year. Three additional clusters ripen progressively later in the spring season within narrowing time windows. We test machine learning methods for predicting the major spring ripening event at each location, based solely on snowpack state. The predictability is proportional to the magnitude of REH, with runoff activation of the highest REH locations predictable within an 18-day window eight weeks in advance.