| dc.description.abstract | Landslide phenomenon continues to be one of the worst natural disasters around the world. Landslide hazard mapping is usually performed through the identification and analysis of hillslope instability factors, recently managed as thematic data within geographic information systems (GIS) environment. Therefore, landslide hazard assessment is normally conducted by analyzing historical landslide events which can be caused by different triggers, including heavy rainfall, earthquake and human activities. In many regions, however, the presence of landslides is not subject to single trigger or event, leading to uncertainties and limitations of single event-based landslide analysis. In recent years, Artificial Neural Network (ANN) is considered as one of the commonly used not only because it can deal with complex and non-linear relationships between slope stability and conditioning factors to predict landslide but also minimize subjectivity. Different from previous studies which mainly focuse on single event-based landslide analysis, for landslide modeling, this study examines multiple landslide events using an ANN approach. Yushan National Park (YNP), located in central Taiwan, was chosen as an ideal test site, as it is a good representative of many geoenvironmental settings within this region and has been continuously affected by typhoons with different magnitude, creating a complex landslide environment. In this research three typhoons that struck Taiwan in 1996, 2001 and 2009 were collected to develop the multiple event-based landslide model using ANN. To develop the ANN model, several landslide occurrence factors were applied including: elevation, slope, aspect, curvature, topographic wetness index, distance to fault, pre-event Normalized Difference Vegetation Index (NDVI), and rainfall variables such as, total rainfall, rainfall intensity, rainfall duration.
Specifically, two experiments were designed to test the applicability of ANN model: (1) single event-based modeling (SEB): developing a landslide model using data from one typhoon event, and validate the model with other two typhoon events; (2) multiple event-based modeling (MEB): developing a landslide model using data from two events and validate the model with the other event. In addition, the ANN model was compared with a linear event-based landslide model, the integrated model (IM), proposed by Chang and Chiang (2009).
To evaluate the model performance, this study used the Receiver Operating Characteristic (ROC) curve, which is based on the proportions of incidences correctly reported as positive (true positive) and incidences erroneously reported as positive (false positive). The area under the ROC curve (AUC) then measures the fitness of the model: the larger the area, the better the model.
The results show multiple event-based models perform better than single ones. Meanwhile, the ANN generally performs better than IM. For ANN-SEB, the validation accuracies vary from 68% – 88% and 87% – 91% for ANN-MEB. For IM-SEB, the validation accuracies vary from 65% – 67% and 75% – 90% for IM-MEB. In average, the validation set accuracies are: 78% for ANN-SEB and 89% for ANN-MEB; 66% for IM-SEB and 84% for IM-MEB. This study demonstrates the importance of considering multiple events in landslide modeling, and also reveals the applicability of ANN method in landslide hazard assessment for Yushan National Park. Finally, this work can be used as a reference to assist slope failure, slope management and tourism planning considering landslide hazard in Yushan National Park. | en_US |