Abstract

Landslides are often catastrophic, causing loss of life and destruction of property and infrastructure. Landslide susceptibility and hazard modeling help mitigate these losses by finding regions prone to landslides and providing probabilistic forecasting of landslide occurrence in a region. However, these landslide susceptibility and hazard models' efficacy depends on the quality of existing databases that often lack crucial information, like the underlying trigger and failure mechanism of a landslide. In this Ph.D. project, we developed methods to identify landslide triggering and failure mechanism information using their geometric and topological properties. For identifying landslide trigger information, we developed three different methods based solely on landslide polygon shapefiles containing landslides' two-dimensional (2D) polygon shapes. The first method uses geometric properties of landslide polygons as a feature space for a machine learning classifier--random forest. In the second method, we transformed these 2D shapes into three-dimensional (3D) point clouds by incorporating the digital elevation data and then extracting landslides' topological properties by topological data analysis (TDA) of these 3D points clouds; and to classify landslides; we treat these topological properties as the feature space of a random forest classifier. The third method uses images of landslides as input to a convolutional neural network (CNN). We tested all three methods using two different testing schemes on six known trigger inventories spread over the Japanese archipelago. In the first scheme, we combine all inventories and then split the dataset into various combinations of training and testing. We train the method on five known triggered inventories in the second scheme and test it on the sixth inventory. Moreover, we implemented each method on an inventory without triggering information to showcase a possible real-world application. The TDA-based method is consistently the most accurate in the above analyses, ranging between 84% to 98% accuracy. To determine the failure mechanism, we explored various geometric and topological properties of landslide shape and found that topological properties are excellent predictors for identifying landslide failure mechanisms. Therefore, we developed a method for determining landslide failure types using landslide topology. First, we extracted the topological features of the landslide 3D shape using Topological Data Analysis and then fed these features as an input to the machine learning algorithm--random forest. We implemented the developed method on the Italian and the US data separately. The method achieves above 95 and 80 accuracies for each landslide failure type for the Italian and US data sets. The methods presented in this Ph.D. dissertation show strong performance in identifying landslide triggers and failure mechanisms. The methods are easy to use as they depend on landslide polygon as input and are transferrable to different regions of the world with adequate training data from areas with similar tectonic and climatic properties. We anticipate that the landslide community and modelers will find our method useful in determining landslides' trigger and failure mechanism. Moreover, we expect that the developed method will enhance the efficiency of landslide predictive models, such as landslide susceptibility and hazards models.

Library of Congress Subject Headings

Landslide hazard analysis; Landslides--Mathematical models; Topology; Machine learning; Landslides--Imaging

Publication Date

3-20-2023

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

Nishant Malik

Advisor/Committee Member

John Kerekes

Advisor/Committee Member

Anthony Vodacek

Campus

RIT – Main Campus

Plan Codes

IMGS-PHD

Share

COinS