Comparison of machine learning models for gully erosion
susceptibility mapping
-
Abstract
Gully erosion is a disruptive phenomenon which extensively affects the Iranian territory, especially in the
Northern provinces. A number of studies have been recently undertaken to study this process and to predict it over
space and ultimately, in a broader national effort, to limit its negative effects on local communities. We focused on
the Bastam watershed where 9.3% of its surface is currently affected by gullying. Machine learning algorithms are
currently under the magnifying glass across the geomorphological community for their high predictive ability.
However, unlike the bivariate statistical models, their structure does not provide intuitive and quantifiable
measures of environmental preconditioning factors. To cope with such weakness, we interpret preconditioning
causes on the basis of a bivariate approach namely, Index of Entropy. And, we performed the susceptibility
mapping procedure by testing three extensions of a decision tree model namely, Alternating Decision Tree
(ADTree), Naïve-Bayes tree (NBTree), and Logistic Model Tree (LMT). We dichotomized the gully information
over space into gully presence/absence conditions, which we further explored in their calibration and validation
stages. Being the presence/absence information and associated factors identical, the resulting differences are only
due to the algorithmic structures of the three models we chose. Such differences are not significant in terms of
performances; in fact, the three models produce outstanding predictive AUC measures (ADTree ¼ 0.922; NBTree
¼ 0.939; LMT ¼ 0.944). However, the associated mapping results depict very different patterns where only the
LMT is associated with reasonable susceptibility patterns. This is a strong indication of what model combines best
performance and mapping for any natural hazard – oriented application.
-
-