| dc.description.abstract | This study uses data set from the “Geological Investigation, Landslide-Debris Flow Investigation and their Susceptibility Evaluation on Watershed” project of the Central Geological Survey, Taiwan (CGS), and aims to improve the susceptibility model developed in the project, by using different rainfall factors and sampling schemes. All the causative factors used in the project are selected in this study; including lithology, NDVI before Aere event, slope roughness, tangential curvature, total slope height, relative slope height, wetness index and fault distance. The final selection of trigger factors is maximum rainfall intensity and total rainfall of the Aere event. Landslide inventories interpreted from SOPT images before and after Aere typhoon event and those of the Matsa typhoon event, were checked by examining rectified aerial photographs, topographic maps, and in the field, so as to establish the event-based landslide inventories. The Aere inventory is used for establishment of susceptibility model and the Matsa inventory for validation. We use different sampling schemes in the study and choose logistic regression as the main analytical method to establish the susceptibility model. We wish to select a set of sample that can expand the range of rainfall values and also raise the weights of the trigger factors. An extra data set from landslide group is selected and is put into the non-landslide group try setting the rainfall value to critical rainfall.
Besides the use of logistic regression for establish susceptibility model, we also adopt discriminant analysis and fuzzy neural network in the study. The results show that AUCs of the success rate curves for logistic regression, discriminant analysis, and fuzzy neural network are 0.8579, 0.8257 and 0.8771, respectively; all show good performance. The results are validated by the data set from of the Matsa typhoon event. The results of validation show that AUCs of the prediction rate curves are 0.7867, 0.7264 and 0.7726, respectively; it is also satisfactory. Although the fuzzy neural network has the highest AUC in establishing model, it is not the highest in validation; this phenomenon may be a kind of over-training. The discriminant analysis obtained the lowest AUC in establishing model and validation. The result of logistic regression is not the best one in establishing the model, but it is the best in validation. Therefore, the logistic regression is the most effective and stable method in establishing the regional landslide prediction model.
| en_US |