Flood Estimation Using Artificial Neural Network Based on Physiographic Features of Watersheds

• Main Authors: Shih-Chien Chan, 詹仕堅
• Other Authors: Chin-Hong Sun
• Format: Others
• Language: zh-TW
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/75067311006636424191

博士 === 國立臺灣大學 === 地理環境資源學研究所 === 91 === Although artificial neural networks have been applied successfully on rainfall-runoff modeling in recent years, the current models constructed by using neural network can neither predict the peak flow and the peak time of flood in ungauged watersheds nor evaluate the hydrological impacts of land use changes. This study offers a solution to resolve the limitations of the construction methods that establish current models. It is suggested that based on the concept of event characterization, physiographic features, which have been ignored in current neural network models, can be and should be put into the neural network learning mechanism, and, together with hydrological features, would thus enable the neural network models to remedy the limitations mentioned above . Four lumped model prototypes of flood estimation are derived from the data of 292 rainfall-runoff events collected from 61 watersheds in various parts of Taiwan, with data from 243 events obtained from 49 watersheds being used to train three-layer structure of back-propagation neural network , and the others for verification. All of the events were characterized as parameters, both hydrological and physiographic, which resulted in a characterizing case-base. In measuring the physiographic features of watersheds, geographic information systems were applied. Three major findings are located in this study. First, the depth of peak flow and the peak time are more efficient estimation targets than peak flow and lag time. Second, the accuracy of the model prototypes adopted in this study is parallel to that of the unit-hydrograph based models. Third, through the implementation of three simulated scenarios of land use change, estimation of hydrological changes rest upon the model prototypes proposed in this study do conform to the principles of hydrology. The general conclusion of this study is that, unlike current neural network models which are inapplicable to flood estimation in ungauged watersheds and the evaluation of the impacts of land use change, the model prototypes presented in this study are free from such deficiencies.