A Machine Learning Approach for Automated Geomorphic Map Generation

Intelligent, automated analysis of data is a critical task in modern data-intensive sciences. The work presented in this thesis is a contribution to this body of research, focusing on automatic geomorphic characterization of planetary surfaces (particularly Mars). We present a framework for automated generation of geomorphic maps from topographic data. Our approach first segments the topographic data into objects and then applies supervised learning for classifying the objects into corresponding geomorphic classes. Choice of optimal segmentation strategy (divisive vs agglomerative) and optimal segmentation resolution are investigated. A custom clustering based agglomerative segmentation algorithm has been developed and implemented, for comparison with an H-Transform based watershed segmentation strategy. Three supervised learning algorithms Naive Bayes, Bagging with decision trees and Support Vector Machines are compared and their map producing abilities evaluated. Finally, a rule-based post processing algorithm has been developed for refining the produced maps. Our algorithm has been tested on five Martian sites and produces encouraging results. The agglomerative segmentation strategy outperforms the conventional watershed based strategy. Furthermore, over-segmented images were found to be more suitable for map generation. Finally, SVM with a quadratic kernel and Bagging with C4.5, both produce accurate maps. Maps produced by SVMs are more robust and visually appealing.