Analysis of the spatial distribution of typhoon-induced rainfall in Taiwan mountains using Gaussian processes

• Main Authors: Ting-Jung Chen, 陳亭蓉
• Other Authors: Hung-Jiun Liao
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/j37sse

碩士 === 國立臺灣科技大學 === 營建工程系 === 94 === This research proposes a probabilistic method of interpolating rainfall amount in the central mountain region of Taiwan during typhoons. The methodology is based on Gaussian process models and the Bayesian analysis. Given data recorded at sparse rainfall stations, the purpose is to estimate the probability density function of the rainfall at any location where rainfall data is absent. Three categories of features are considered to control the rainfall variability, including topographical features, GIS features, and typhoon-related features. It is found that topographical features are more important than the others. In particular, among all the seven features, longitude and latitude of the location of interest are usually the two most important features; the GIS-based average elevation is more influential than the point-wise elevation. It is also found that the percentage of the prediction errors reduces with increasing rainfall and that the interpolated rainfall in the central mountain may increase or decrease with elevation depending the characteristics of the typhoon. The analysis results of the Gaussian process model are compared with those obtained by the inverse-distance-weighting (IDW) method and by the Kriging method. The prediction errors for the three methods are comparable; however, the IDW and Kriging methods only provide point estimates, while the Gaussian process method can also provide the probability density function of the predicted rainfall. This additional information of probability distribution is valuable since there are lots of uncertainties in the interpolation process; therefore, it is essential to quantify the associated uncertainties.