Spatial prediction of landslide hazards and risk areas using integrated statistical and data mining approaches

Landslides are one of the many forms of natural hazards that often cause severe property damages, economic loss, and high maintenance costs. Slope failures are a result of multiple triggering factors, including anthropogenic activities, earthquakes,and intense rainfall, and of reactions of a host of unstable surface materials related to geology, land cover, slope geometry, moisture content, and vegetation. This thesis presents a set of novel GIS-based statistical approaches developed for the hazard mapping of rainfall-induced landslides. These approaches were tested in two study areas: (1) Kuala Lumpur (KL) and surrounding areas in Malaysia and (2) Pohang– Gyeongju area in South Korea. This research has four objectives; the objective number one focuses on developing an alternative technique for landslide inventory modeling based on spatial pattern characterization. The cluster patterns of inventory were extracted and used as training data in constructing the landslide susceptibility models. The resultant map showed that the prediction results of the proposed approach were more accurate than those of the random selection method. The results also exhibited a noticeable reduction in uncertainty, particularly in landslide-prone areas with high to moderate hazard. The objective number two focuses on the statistical correlation between landslide conditioning factors and the location of occurrence. The ensemble methodology was employed for pairwise relationships generated among the spatial factors by integrating the weights of the evidential belief function (EBF) into the analytical hierarchy process (AHP). The results successfully determined the conditioning factors most relevant to landslide occurrence. The interrelationship between the conditioning factors with respect to landslide occurrence was also considered in the analysis. The objective number two also seeks to develop an ensemble methodology that uses chi-squared automatic interaction detection (CHAID) for the autoclassification of the landslide conditioning factors and whose results are integrated into the logistic regression (LR) model. The most significant landslide conditioning factors were prioritized, and the developed models were validated using success and prediction rate curves. The research findings confirmed the research hypothesis by testing the proposed hybrid ensemble models with landslide locations which were not used during model building, and comparing it with existing individual models. The prediction maps yielded higher prediction accuracy and achieved better discrimination of susceptible zones than the other models did. The objective number three focuses on conducting rainfall threshold analysis on Kuala Lumpur (KL) and the surrounding areas. Four different antecedent periods: 5-,10-, 15-, and 30-day relationships, were used in defining the rainfall threshold of each rainfall station in the study area. The results fill a gap in the literature through the formation of a medium-scale hazard map that was developed based on the multiplied results of the spatial and temporal landslide susceptibility maps of KL and the surrounding areas using available information from 2000 to 2012. The objective number four focuses on the semi-quantitative risk assessment of landslide hazards in the western and northern regions of KL; only medium-scale data were used because of data scarcity. A valid integration between the elements at risk and the hazard map of 2017 was then accomplished to predict the number of elements that are likely to be affected by direct risks. The resultant methodology employs an exposure-based analysis method to calculate the number of elements at risk and prioritizes land use, population, and road networks. Results showed the approximate percentage of loss in the following areas: about 50% for residential dwellings, 35% for commercial buildings, 17% for industrial buildings, 31% for utilities, 18% for the population, and 27% for road networks. The results prove the capacity of the proposed method to make valid predictions under landslide risk conditions in a data-scarce environment. Missing attributes from damaged records rendered the validation of current findings an impossible task. The results are expected to not only provide a quick yet comprehensive assessment of future landslide hazards and risks but also serve as a guide for land use planners. The presented methods and information will add a valuable contribution to the landslide hazard and risk assessment of medium scale data analysis.

Preview PDF FK 2014 21R.pdf Download (1MB) | Preview

Actions (login required)

View Item