Coupling GIS and Machine Learning for the quantitative assessment of geomorphodiversity

Abstract

<p>The aim of this work is to use machine learning methods coupled with GIS to evaluate the Geomorphodiversity Index (GmI).</p><p>The starting assumptions are that:</p><p>-          the geomorphodiversity or the variety of landforms reflects not only the geomorphological processes but the geological component too.</p><p>-          The numerical assessment is preferable to have a spatial distribution of the geomorphodiversity, being an objective and repeatable procedure and allowing for the comparison of areas in different geographical contexts.</p><p>-          The digital data and in particular the Digital Elevation Models, can define the topographic attributes necessary to gain the quantitative index without taking into account the traditional geomorphological maps. The topographic attributes reveal and summarize the presence and the efficiency of the driving forces that shape the Earth surface. The drainage density could fill the gap on the flat areas where the topographic attributes failed (for medium horizontal resolution values).</p><p>Trying to do an unsupervised clustering approach a K-Means and Mini-Batch K-Means for "partitional" clustering and Agglomerative Clustering for "hierarchical" clustering algorithms have been used with Anaconda (an open-source software distribution platform used for data analysis and specifically, Spyder was used as IDE -Integrated Development Environment- and Python 3.7 as a programming language) coupled with ArcGIS 10.1 © ESRI software. The research is split in three different steps: the selection of the data into ArcGIS, managing, analysis and clustering of the dataset with Python, reclassification of the data and comparison with the GmI already known in the literature in ArcGIS.</p><p>The test area is the Umbria region (central Italy) where the GmI derived from a simple GIS analysis is already available.</p><p>The advantages of the unsupervised method are that:</p><p>-          no weight was previously assigned to the individual typographic parameters.</p><p>-          Calculation times are extremely short.</p><p>-          The results are similar but more accurate than GIS analysis alone.</p>