Technical Report - Methods: Automated Discovery of Functional
Relationships in Earth Systems Data
Abstract
Functional relationships capture how variables co-vary across specific
spatial or temporal
domains. However, these relationships often take complex forms beyond
linear, and they may
only hold for sub-sets of the domain. More problematically, it is often
a priori unknown how
such sub-domains are defined. Here we present a new method called SONAR
(diScovery Of
fuNctionaAl Relationships) that enables the automated discovery of
functional relationships in
large datasets. SONAR operates on existing unstructured data and is
designed to be an
explorative tool for large datasets where manual search for functional
relationships would be
impossible. We test the method on groundwater recharge outputs of
several global hydrological
models to explore its usefulness and limitations. Further, we compare
SONAR to the established
CART (Classification and Regression Trees) and CIT (Conditional
Inference Trees) methods.
SONAR results in smaller trees with functional relationships in the leaf
nodes instead of specific
classes or numbers. SONAR provides a robust and automated method for the
exploration of
functional relationships.