宜蘭地面降雨之統計降尺度推估：以IPCC-SRA2情境為例

Statistical Downscaling of Rainfall in Yi-Lan Region: Case study of IPCC-SRA2 scenario

王振剛(Cheng-Kan Wang)

指導教授 ： 余化龍

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/碩士(2011年)

https://doi.org/10.6342/NTU.2011.01995

摘要

本研究之主要目的為建立臺灣東北部宜蘭地區地面降雨之統計降尺度模式，並以統計降尺度方法推估此地區之單點月降雨與區域週降雨，利用支援向量機(Support Vector Machine, SVM)的非線性回歸能力建立大尺度大氣網格變數與地面測站觀測降雨間之非線性關係。
　　由於臺灣地區降雨機制極為複雜，在各季節中有相當明顯的不同，而傳統的春、夏、秋、冬四季無法妥善地區分這些降雨機制，因此本研究中結合EOF與K-means聚類方法，依據周圍區域之大氣、降雨特徵，將臺灣地區之降雨型態重新分類為四個降雨季節，分別為春、梅雨季(四至六月)、颱風季(七至八月)、季風轉換季(九至十一月)及東北季風季(十二至三月)，並對四個降雨季節分別建立其降雨降尺度模式。
　　在模式建立的過程中，由於大氣變數之資料維度極高，本研究使用經驗正交函數(Empirical Orthogonal Function, EOF)方法萃取各大氣變數之特徵，以降低資料維度，利於模式之建立與訓練，同時以真實的大氣系統解釋萃取出之大氣特徵，試圖在統計降尺度方法中加入物理依據。
　　最後的結果顯示，本研究方法所建立之降尺度模式在春、梅雨季中獲得較為良好之結果，而在颱風季與季風轉換季表現較差。應用模式於GCM未來資料的推估，顯示未來於梅雨季的降雨將有顯著的增加，颱風季中雨量則略為減少，但並不顯著，九至十一月由本研究模式模擬之結果雖不完全準確，但亦可指示整體雨量當有明顯地增加，冬季的雨量則幾乎無變化。

關鍵字

統計降尺度 ； 降雨 ； 支援向量機 ； 經驗正交函數 ； K-means

並列摘要

The purpose of this study is to develop a statistical downscaling model to downscale the monthly and weekly rainfall in Yi-Lan region, which located at northeastern Taiwan.
This study included two parts, single point downscaling and regional downscaling, monthly rainfall observation from only Yi-Lan weather station was used for the former, and weekly rainfall observation from 24 rainfall stations located at Yi-Lan region were used for the latter. The space-time variations of climate variables and monsoon features among the season lead to distinct local rainfall patterns in study area. Hence, this study combined EOF method and K-means clustering to classify a year into four seasons, which were mei-yu season (April to June), typhoon season (July to August), transition season (September to November), and northeastern monsoon season (December to March). Downscaling model was then established base on this classification.
To establish the nonlinear relationship between large scale climate variables and the local regional rainfall, Support Vector Machine (SVM) and Empirical Orthogonal Function (EOF) method were mainly used. The EOF method was used not only to reduce the data dimension of the space-time climate variables and local region rainfall observation, but to identify the most important spatial patterns of each climate variables and local rainfall during the study period. This approach is expected to add some real physical meaning in statistical downscaling method.
In results, the model performed well in only mei-yu season, but not quite well in typhoon and transition seasons. Also, in the GCM future scenario, it shows that the rainfall significant increase in mei-yu and transition season, decrease in typhoon and northeastern monsoon season, but not significant. Though the model performance in transition season is not quite well, the model could still indicate that trend of rainfall in this season will apparently increase.

並列關鍵字

statistical downscaling ； rainfall ； support vector machine ； empirical orthogonal function ； K-means

參考文獻

朱錦紅, 柯., 許晃雄 (2002). "臺灣降水變化研究回顧-颱風長期變化與降水長期預報." 全球變遷通訊雜誌 36.

朱錦紅, 柯., 許晃雄 (2002). "臺灣降水變化研究回顧— 氣候分區、降水與東亞大氣環流之關聯." 全球變遷通訊雜誌 12: 28-37.

Biau, G., E. Zorita, et al. (1999). "Estimation of precipitation by kriging in the EOF space of the sea level pressure field." Journal of Climate 12(4): 1070-1085.

Braud, I. and C. Obled (1991). "On the Use of Empirical Orthogonal Function (Eof) Analysis in the Simulation of Random-Fields." Stochastic Hydrology and Hydraulics 5(2): 125-134.

Brian D. Ripley, N. L. H. (1995). Pattern Recognition and Neural Networks, Cambridge University Press