| Summary: | 碩士 === 長榮大學 === 土地管理與開發研究所 === 94 === Owing to the steep morphology, young and weak geological environment, heavy rainfall and inappropriate land use, Taiwan is subject to occurrence of debris flow, which often causes enormous human and material losses. Hence, the forecasting of debris flow is important to provide advance warning of debris flow. Due to lacking of enough disaster occurrence records to setup the warning criteria for each debris flow gully, the objective of the study is to develop the evaluation model on assessment of occurrence frequency of debris flow hazardous gullies.
In the present study, the eight factors related to the occurrence of debris flow including rainfall intensity, watershed area, average slope of watershed and gullies, shape factor of watershed, landslide ratio, geological patterns and land-use were adopted to clustering 413 debris flow hazardous gullies in the Central Taiwan. The variance between different clustering groups was verified through Wilk’s value and disaster occurrence records. Clustering results show that 413 debris flow hazardous gullies in the central Taiwan can be reasonably divided into 3 groups, and the average occurrence frequency in three main groups is 0.97, 0.71 and 0.56, respectively. The occurrence frequency was then estimated according to the membership function of each gully belong to three main groups. The numbers of debris flow occurrence usually greater than 3 when frequency exceed 0.7. The numbers of debris flow occurrence always less than 2 when frequency under 0.6.
Hence, artificial neural networks method was applied to develop the evaluation model on assessment of occurrence frequency of debris flow hazardous gullies. 331 debris flow hazardous gullies were used to training sample and 82 gullies were used to test sample in neural network. Based on the different algorithms, two models were tested and compared. The result shows that the Liebenberg-Marquardt (LM) method with two hidden layers, and three neurons in each hidden layer, has better performance in occurrence frequency forecasting. The value of mean absolute percentage error (MAPE) and correlation coefficient (r2) is 2.11% and 0.97, respectively.
|
|---|
