Radar Precipitation Estimation Using Support Vector Machine

• Main Authors: Huang, Bo-Jhen, 黃柏禎
• Other Authors: Tsai, Chun-Ming
• Format: Others
• Language: zh-TW
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/60276377877488607102

碩士 === 臺北市立大學 === 資訊科學系碩士在職專班 === 103 === Significant problems exist for traditional precipitation estimates from radar when it comes to comparing with rain gauges on the ground. To avoid the problem and better utilize the information in the four-dimensional structure of the atmosphere, this research proposes a precipitation estimation method based on support vector machine and regression in the hopes to improve the accuracy of precipitation estimates. To generate various feature sets as inputs for the experiments, a set of tools have been developed, including a radar feature extraction and labeling tool based on AWIPS II (Advanced Weather Interactive Processing System) and a library for further batched processing of the extracted features such as calculation of statistics and combinations of features from various elevations and/or products. For experiments of this research, Taiwan Central Weather Bureau’s Wu-Fen-Shan weather radar products, including reflectivity (Z), differential reflectivity (ZDR) and specific differential phase (KDP) from rainy days in the years of 2012, 2013 and 2014 are used as the source of features, and the target location for precipitation estimation is Taipei weather station. In one of the experiments, we first aggregated over the feature vectors within the same observation hour to derive a mean feature vector, then partitioned 10-minute precipitation accumulations into 6 classes as labels of the features, by using the resulting feature set along with support vector regression, the best result gives a root mean squared error of 0.54 and correlation coefficient of 0.95, corresponding to features derived from 0.5-degree elevation specific differential phase (KDP), when the extreme 10-minute precipitation value of 9.4(mm) in the test set is removed, the corresponding root mean squared error becomes 0.5 and correlation coefficient dropped to 0.72. Results from this reaearch also indicate the proposed support vector regression estimation method has about the same performance as traditional R(KDP) given by Sachidananda and Zrnic(1987) and is better than R(ZDR,KDP) by Ryzhkov and Zrnic(1995). In summary, using combinations of statistics such as middle value, mean, and maximum 5 values calculated from 5x5 feature vectors of low-elevation KDP products, along with support vector regression, is the best solution found for radar precipitation estimation in this research. While combinations of Z、ZDR and KDP as well as combinations of various elevations are also tried, no significant improvements can be derived in this research.