結合多維度時空因子與都卜勒雷達觀測於區域雨量推估

Abstract

現今全球都因氣候異常而面臨許多災害問題，而這些災害中又以與降雨有關的災害佔最大比例。世界各國為了降低這些因降雨而導致的損失以及增加面對災害時的應變能力，各自都有不同的作為。其中又以成本效益高、施行時間快速的非結構性措施最受各國青睞。為了使非結構性措施能夠有效運用於防災救災，獲取各地正確的降雨是至關重要的依據。目前獲取雨量資訊的主要工具分為雨量站及雷達回波，雨量站能夠直接蒐集雨量資訊，提供準確的點狀資料；雷達回波則是採用間接觀測的方式，將蒐集到的反射率轉換為降雨率，提供具參考性的面狀資料。本研究將利用上述之蒐集與量工具並考慮了其他可能對降雨會產生影響的因子，透過深度學習的方式建立較高準確度的雨量推估模型，也將討論時間序列的模型對於降雨量的推估是否會優於非時間序列模型，透過分別建立非時間序列模型多層感知器(Multi – Layer Perceptron, MLP)及時間序列模型長短期記憶(Long Short – Term Memory, LSTM)，得出在記憶時間合宜的情況下，時間序列模型能有更好的結果。針對合宜的時間模擬了18個小時內的情境，得到只記憶4個小時以內的結果與非時間序列模型的結果無顯著上的差異，透過因子重要性排序發現原因為先驗值的貢獻程度低，導致無法發揮時間序列模型的優勢；而記憶至少14個小時的結果則為最佳。另外根據因子重要性排序得到，影響雨量值的主要因素包含經緯度及月份，透過與中央氣象局發行的氣候年報比對，證實了降雨與上述因子確實有著相當的關連性。經上述研究證實，時間序列模型相較於非時間序列模型，對於大雨等級以下的雨量推估有較好的結果，此外將影響降雨的因子納入模型中也能幫助獲得更準確的資訊。

Nowadays, the world is facing numerous disaster issues due to abnormal climate conditions, with rainfall-related disasters comprising the largest proportion among them. Countries around the world have taken various measures to reduce losses caused by rainfall and enhance their resilience in the face of disasters. Nonstructural measures that are cost-effective and can be implemented quickly have gained popularity among countries. To effectively utilize nonstructural measures for disaster prevention and mitigation, obtaining accurate rainfall data is crucial. Currently, the primary tools for obtaining rainfall information are rain gauges and radar echoes. Rain gauges directly collect rainfall information and provide accurate point data; radar echoes, on the other hand, employ an indirect observation method by converting collected reflectivity into rainfall rates, providing informative spatial data for reference. Besides two conventional tools that collect rainfall values, this study also considers other factors that may have an impact on rainfall. This study also combines accurate rainfall station data, informative radar echoes, and multiple spatial-temporal factors affecting rainfall to establish a high-accuracy rainfall estimation model using deep learning techniques. Investigating whether time series models outperform non-time series models in rainfall estimation. Comparion is conducted by separately modeling a non-time series model MLP and a time series model LSTM. Concluded that the time series model demonstrates better performance in rainfall estimation. Furthermore, incorporating the factors influencing rainfall into the model also contributes to obtaining more accurate information.