GEOMORPHOLOGIC CLASSIFICATION OF COLOMBIA BY MACHINE LEARNING WITH DIGITAL ELEVATION MODEL

摘要

There are many natural disasters in Colombia and seismic hazards depend on the terrain, therefore hazard maps are required. Geomorphologic classification maps are often used for baseline data to create hazard maps; however, conventional geomorphologic classification methods by visual interpretation require expertise and time, and no terrain classification map exists for the whole extent of Colombia. An automatic geomorphologic classification method produced from a digital elevation model solves these problems, but the classification result depends on the region of interest, and accuracy in Columbia is insufficient due to lack of features. To solve these problems, we developed a practical and labor-saving geomorphologic classification method by machine learning with an existing geomorphologic classification map as supervisor and created a new classification map for all of Colombia. Existing detailed geomorphologic classifications in some areas were integrated into 17 classes based on the classification criteria of Japan's engineering-based geomorphologic classification map for natural hazard evaluation. In addition to slope (gradient), texture (ridge/valley density) and convexity (convex distribution density) of the previous study, the features of elevation value, texture continuous value (ridge/valley density calculated by continuous value), convexity continuous value (convex density calculated by continuous value), distance from coastline, distance from river, relative relief from the nearest river and the RGB values of the optical satellite image were newly introduced. The three machine learning algorithms of Random forest, Neural network and SVM were carried out using these feature values, resulting in approximately 77% overall accuracy for Random forest. Accurately classified geomorphologic classes tended to differ from other classes in partial dependent plots and the distribution of specific feature values. Finally, using a classifier generated by Random forest, a geomorphologic classification map across the whole of Colombia was created in combination with existing geomorphologic classification maps.