Using Machine Learning to identify drivers of heterogeneous behavior in Greenland outlet glaciers

Nearly half of Greenland's interior ice is transported through outlet glaciers to the ocean, yet their response to environmental and local parameters is poorly understood. The dynamics of outlet glaciers are influenced by a wide range of external climatic properties such as atmospheric and ocean forcings, as well as local properties such as bed topography and subglacial hydrology. Heterogeneity in glacier dynamics has been demonstrated in previous studies, however, a robust quantitative analysis of temporal glacier dynamic changes and their correlations to external forcings has not been holistically investigated. By extending the time series of glacier dynamic variables (terminus location, velocity, and surface elevation) and environmental forcings (e.g., ocean thermal forcing, surface mass balance), we use machine learning to cluster the behavior of a set of west Greenland outlet glaciers. Additionally, using a long short-term memory neural network (LSTM), we evaluate the statistical correlations between observed glacier dynamic behaviors and environmental forcings within each cluster and identify the environmental forcings that have the greatest impact on the dynamics of each cluster. These findings will help improve calibration of physically based numerical models and improve sea-level rise estimates.