Application of Neural Network to Landslide Susceptibility Analysis
| Autor: | Yen-Hsiang Lin, 林彥享 |
|---|---|
| Rok vydání: | 2003 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 91 There is no common agreement on the question, that is, which is the best method among various landslide susceptibility analyses, since they all have their own advantages and disadvantages. Our research tried to utilize the self-learning capability of back propagation neural network and to establish the relationship between landslides and the factors that potentially affecting landslides. Landslide susceptibility index of each cell of study area were then calculated by the recalling process of neural network to prepare the final landslide susceptibility map. The study area of 23km in width and 30km in length is located in Central Taiwan and between the Da-Li town and the Kuo-Shing town. Landslide inventory has been mapped and digitized from SPOT satellite images prior and after the Chi-Chi earthquake. Landslide inventory based on aerial photos from ERL/ITRI has been used for comparison and validation. The 1/5,000 topographic maps published after the Chi-Chi earthquake were also used for validation. Landslide which is too small or questionable was removed during this process. Moreover, the deposit part at the toe of landslide which will confuse the neural network leaning has also being removed. Erdas Imagine and Mapinfo were used for spatial data analysis in this study. We also developed some Fortran programs for extracting and processing the landslide susceptibility factors. Bivariate statistical method was used to select the relative important factors for the input layer of neural network. We classified our training data into three groups, the first group is the stable area which has slope less than 10%, the second group is the non-landslide pixels, and the third ones is the earthquake triggered landslide pixels. Concept of fuzzy logic was then adopted to design a membership function for each group and the membership function accepted as the output layer of neural network. The results of this study are: (1) The combination of one-hidden-layer and 18 neuron have resulted in the highest accuracy rate. (2) Two-hidden-layer has better convergence in erron function than one-hidden-layer, but it is worse in accuracy rate than one one-hidden-layer. This implies that the two-hidden-layer setting may have over-complicated the problem. (3) The best accuracy rate for predicting the landslide using the neural network is 93.7%. (4) Landside susceptibility value can be calculated from fuzzy membership function, and then be used to construct a map. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|