Marwah Al Ismail Dissertation Defense
Advances in Methodology for Regional Landslide Analysis using Statistical, Mechanical, and Machine Learning Approaches
Landslides are catastrophic events that occur on critically unstable slopes in steep, mountainous regions in response to external triggers, such as seismic events or prolonged,
intense rainfall. The damage resulting from those landslides is mostly determined by their magnitude and size. Consequently, assessing landslide hazards in steep, tectonically active mountains at high altitudes remains a challenging task; however, advances in computational methods and technologies can make it achievable. This dissertation demonstrates the novel joint utility of mechanistic and statistical (chapters 2 and 3) and machine-learning (chapter 4) techniques for landslide analysis to understand the contribution of near-subsurface material strength, topography, and triggering factors to the regional distribution of landslide sizes.
In Chapters 2 and 3, we use computational methods to assess hillslope stability using geotechnical engineering models, combined with extensive probabilistic topographic
segmentation, to generate a synthetic distribution of landslide sizes. Specifically, we use 2D slope stability analysis to assess the probability of failure (P f ) for a given hillslope geometry (i.e., hillslope area and gradient) under subsurface strength conditions (i.e., cohesion and angle of friction). In addition, using hydrological tools to discretize topography in a random, probabilistic approach identifies the distribution of hillslope geometries prone to failure. The combined probabilities from mechanistic modeling and the availability of hillslopes in a given topography result in a synthetic distribution of landslide sizes. This modeled distribution allows us to investigate, in great detail, the impact of the variation in strength–and hence weathering–with depth and slope, as well as the topography on the overall distribution. Using two different landslide inventory datasets: the 1994 Mw6.7 Northridge earthquake in Southern California and the 2015 Mw7.8 Gorkha earthquake in central Nepal, we find that the variability in the weathering gradient with depth and the slope gradient is essential for accurate reproduction of regional landslide size distribution. Moreover, because of the variable topographic gradient and climatic conditions within central Nepal, we apply this framework to back-estimate the strength-depth profile for each sub-region within the study area from south to north. We find that back-estimated strength increases northward, coinciding with increases in the topographic gradient and landslide density. These findings demonstrate the significance of the synthetic
distributions for identifying the controls on the regional landslide size distributions as well as providing an insight into the critical zone structure for hazard assessment and landscape evolution studies.
Chapter 4 provides an opportunity to explore the use of machine learning models for landslide analysis, specifically for improving rainfall-triggered landslide forecasting. Steep mountain belts at high altitudes experience the orographic effect, which produces intense rainfall and, consequently, triggers landslides. As rainfall-triggered landslide sizes are sensitive to rainfall event characteristics (i.e., storm depth and intensity), using accurate rainfall data with high spatial and temporal coverage is essential. Ground-based rainfall measurements from gauge stations provide an accurate source of data; however, gauge stations are usually sparsely located and limited in steep terrain, where landslides are most common. In recent years, remotely sensed precipitation measurements have provided high spatial and temporal resolution data, but are often inaccurate, especially on hillslopes at high altitudes. To overcome challenges in each dataset source, we use machine learning models to predict calibrated rainfall data at locations where gauge stations are absent, with a focus on the monsoon season, during which rainfall intensifies. The use of machine learning approaches and associated evaluation metrics enables us to validate their utility for predicting the total rainfall during the monsoon season with high certainty, compared with the original remotely sensed rainfall data. In addition, the model’s predicted rainfall captured the annual distribution of storms during the monsoon season. Although the model’s accuracy at predicting the timing of extreme rainfall events (which trigger landslides), this work provides a promising avenue for using machine learning techniques to predict high-resolution, high-temporal-resolution rainfall datasets.
On a broader scale, this dissertation provides novel methodologies that enable us to improve landslide hazard analysis, assessment, and forecasting.
intense rainfall. The damage resulting from those landslides is mostly determined by their magnitude and size. Consequently, assessing landslide hazards in steep, tectonically active mountains at high altitudes remains a challenging task; however, advances in computational methods and technologies can make it achievable. This dissertation demonstrates the novel joint utility of mechanistic and statistical (chapters 2 and 3) and machine-learning (chapter 4) techniques for landslide analysis to understand the contribution of near-subsurface material strength, topography, and triggering factors to the regional distribution of landslide sizes.
In Chapters 2 and 3, we use computational methods to assess hillslope stability using geotechnical engineering models, combined with extensive probabilistic topographic
segmentation, to generate a synthetic distribution of landslide sizes. Specifically, we use 2D slope stability analysis to assess the probability of failure (P f ) for a given hillslope geometry (i.e., hillslope area and gradient) under subsurface strength conditions (i.e., cohesion and angle of friction). In addition, using hydrological tools to discretize topography in a random, probabilistic approach identifies the distribution of hillslope geometries prone to failure. The combined probabilities from mechanistic modeling and the availability of hillslopes in a given topography result in a synthetic distribution of landslide sizes. This modeled distribution allows us to investigate, in great detail, the impact of the variation in strength–and hence weathering–with depth and slope, as well as the topography on the overall distribution. Using two different landslide inventory datasets: the 1994 Mw6.7 Northridge earthquake in Southern California and the 2015 Mw7.8 Gorkha earthquake in central Nepal, we find that the variability in the weathering gradient with depth and the slope gradient is essential for accurate reproduction of regional landslide size distribution. Moreover, because of the variable topographic gradient and climatic conditions within central Nepal, we apply this framework to back-estimate the strength-depth profile for each sub-region within the study area from south to north. We find that back-estimated strength increases northward, coinciding with increases in the topographic gradient and landslide density. These findings demonstrate the significance of the synthetic
distributions for identifying the controls on the regional landslide size distributions as well as providing an insight into the critical zone structure for hazard assessment and landscape evolution studies.
Chapter 4 provides an opportunity to explore the use of machine learning models for landslide analysis, specifically for improving rainfall-triggered landslide forecasting. Steep mountain belts at high altitudes experience the orographic effect, which produces intense rainfall and, consequently, triggers landslides. As rainfall-triggered landslide sizes are sensitive to rainfall event characteristics (i.e., storm depth and intensity), using accurate rainfall data with high spatial and temporal coverage is essential. Ground-based rainfall measurements from gauge stations provide an accurate source of data; however, gauge stations are usually sparsely located and limited in steep terrain, where landslides are most common. In recent years, remotely sensed precipitation measurements have provided high spatial and temporal resolution data, but are often inaccurate, especially on hillslopes at high altitudes. To overcome challenges in each dataset source, we use machine learning models to predict calibrated rainfall data at locations where gauge stations are absent, with a focus on the monsoon season, during which rainfall intensifies. The use of machine learning approaches and associated evaluation metrics enables us to validate their utility for predicting the total rainfall during the monsoon season with high certainty, compared with the original remotely sensed rainfall data. In addition, the model’s predicted rainfall captured the annual distribution of storms during the monsoon season. Although the model’s accuracy at predicting the timing of extreme rainfall events (which trigger landslides), this work provides a promising avenue for using machine learning techniques to predict high-resolution, high-temporal-resolution rainfall datasets.
On a broader scale, this dissertation provides novel methodologies that enable us to improve landslide hazard analysis, assessment, and forecasting.
| Building: | 1100 North University Building |
|---|---|
| Event Type: | Lecture / Discussion |
| Tags: | Earth And Environmental Sciences |
| Source: | Happening @ Michigan from Earth and Environmental Sciences |
