Automated neural network identification of cirques

Abstract

Morphological characteristics of cirques have been studied for decades; however, no repeatable set of metrics has been derived that can consistently identify them. Perhaps more importantly, there is no consensus definition of the form that distinguishes cirques and clusters of cirques from non-cirques. In our approach, we use Shuttle Radar Topography Mission (SRTM) digital elevation models (DEMs) in a Convolutional Neural Network (CNN) framework to identify cirques in 20 mountain ranges globally. We extracted bounding boxes of cirques in 19 of these study areas and used them to develop a training set for a cirque identification model. The trained model was applied to the Sierra Nevada California to assess whether this algorithmic approach derived from a global dataset could produce consistent results in complex terrain with mutually interacting cirque forms. Using commonalities revealed using this approach, we find that there is a basic, recognizable and morphometrically quantifiable cirque form. This approach can be used to automate the identification of cirque locations and to guide the quantification of cirque form independent of the subjective definitions of individual workers. The approach can also be used to understand cirque form under different environmental conditions, including similar forms on Mars.