請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/53961
標題: | 應用深度學習發展山區水文敏感區之雷達QPE方法 Development of A Deep Learning Radar QPE Technique for Hydrologically Sensitive Mountain Areas |
作者: | Chia-Tung Chang 張佳棟 |
指導教授: | 李清勝(Cheng-Shang Lee) |
關鍵字: | 山區水文敏感區,定量降雨估計,深度學習,卷積神經網路, Hydrologically Sensitive Mountain Areas,Quantitative Precipitation Estimation (QPE),Deep Learning,Convolutional Neural Network (CNN), |
出版年 : | 2021 |
學位: | 碩士 |
摘要: | 本研究利用卷積神經網路模型(convolutional neural network,CNN)、萃取雷達CFAD(Contoured frequency by altitude diagram)中對流垂直結構之特徵,建立該特徵和降雨量之間的非線性關係,並強調時雨量大於10 mm(大雨)之降雨估計準確度。主要目的係利用深度學習演算法能由複雜的多維度、多變數資料、自動分析資料中重要特徵的優點,並結合CFAD中、高層回波資訊,克服傳統定量降雨估計(quantitative precipitation estimation,QPE)技術在山區、最低可用仰角受地形遮蔽或偏高的問題。模型訓練過程中,嘗試在輸出雨量值前,加入ReLU整流函數、量化大氣濕度;並針對降雨資料呈現不均勻分布的問題,為大雨事件加入權重、以平衡資料比例。此外,在模型損失函數加入「20 mm以上之降雨分類」,以提高模型對大雨的掌握能力。 結果顯示,針對山區的大雨事件,CNN模型估計之時雨量的均方根誤差為10.15 mm,較Z-R關係式(Z=〖32.5R〗^1.65)者小0.93 mm;此CNN模型在30~40 mm大雨區間與Z-R關係式相比有最大進步幅度。針對獨立個案瑪利亞颱風(2018)的分析結果顯示,CNN模型在大屯山區域之降雨估計十分準確;而針對平等測站所觀測到的數個降雨峰值(30~35 mm),CNN模型皆能夠合理掌握,誤差不超過10 mm;Z-R關係式則有較大誤差。整體而言,在山區、CNN模型和Z-R關係式對於小雨事件皆有不錯掌握能力,當雨量較大時,則CNN模型有較好的表現。然而,隨著雨量門檻值的增加,二者皆有低估降雨之情形,不過在時雨量小於40 mm 時,CNN模型的估計值仍有相當高的參考價值。意即CNN模型能有效利用對流特徵與降雨間的關係,改善山區QPE技術;但對於極端降雨(≥40 mm)事件的估計,CNN模型則尚有改進空間。 Convolutional neural network (CNN) model is established to extract features from vertical convective structures in radar contoured frequency by altitude diagrams (CFADs). This CNN model is then used to estimate hourly rainfall based on the nonlinear relationship between the CFADs and rainfall; the model emphasizes the accuracy of events with hourly rainfall larger than 10 mm (heavy rain). As the CNN can automatically extract features from high altitude reflectivity information in CFAD, our model solves the partial beam blockage and “lowest available elevation angle too high” problems that traditional QPE (quantitative precipitation estimation) faces in mountain areas. Our model has some unique designs compared to a conventional CNN model, including (i) a ReLU rectification layer is added before the model outputs the rainfall value so that the model can quantify the atmospheric humidity; (ii) to handle the uneven distribution of rainfall data, we add weights to heavy rain events and balance the data ratio; (iii) the 'rainfall classification above 20 mm' is added on the loss function to emphasize the performance in heavy rain. The results show that in the mountains and for heavy rain events, the root mean square error of the CNN model is 10.15mm, which is 0.93mm less than that of the Z-R relation (Z=〖32.5R〗^1.65). Furthermore, the CNN model gets most considerable improvement than Z-R relation in a rainfall range from 30 to 40 mm. For Typhoon Maria (2018), the CNN model is more accurate in the Datun mountain area. Specifically, multiple heavy rainfall peaks were observed at the Pingdeng station, and the CNN model can correctly grasp them with an error less than 10 mm while the Z-R relationship has a larger error. In general, both the CNN model and the ZR relationship have a good mastery for the light rain events in mountain areas; however, CNN model performs better when rainfall becomes larger. Also, for large rainfall thresholds, both of them suffer from the rainfall underestimation. CNN model’s estimation is relatively reliable when hourly rainfall is less than 40 mm. Overall, the CNN model can efficiently improve the QPE technique for heavy rainfall events in mountains; however, for the estimation of extreme rainfall (≥40 mm) events, the CNN model still has room for improvement. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/53961 |
DOI: | 10.6342/NTU202100564 |
全文授權: | 有償授權 |
顯示於系所單位: | 大氣科學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-0502202109170500.pdf 目前未授權公開取用 | 3.74 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。