| 摘要: | 近年來，由於隨著氣候變遷的衝擊與台灣山高水急的自然地形特性，降雨量隨著時空分布十分不均，且人口稠密，農業、工業、經濟發展上導致水資源開發不利、地下水補注量降低，更因時空降雨特性改變所衍生出來的種種問題，包括許多極端水文事件與環境問題等。台灣多處地區，尤為南部地區其主要降雨季節集中於每年5-6月之梅雨季節以及7-9月之颱風季節，豐水期與枯水期之降雨量存在顯著差異，然而台灣用水需求主要仰賴梅雨與颱風所挾帶的降雨，因此降雨季節雨量的些微改變皆會直接或間接地影響台灣面臨旱澇之災害風險，因此了解降雨於各地區時空變異特性及影響其變動的自然因子並加以預測為目前重點關注課題之一。
本研究針對台灣本島，以資料探勘(data mining)的角度利用實際的觀測資料並且結合巨觀時間與空間尺度下的遙測資料，透過小波訊號分析方法進行特徵萃取，探討台灣地區雨量與可能影響水文水資源變動特性因子之間關連性，發現太陽黑子有著顯著之10至12年週期，太陽黑子(SSN)的數量根據大約11年的週期變化；而南方震盪指數(SOI)則是於2至8年有較為顯著的週期相關性，其主要與聖嬰-南方震盪現象(ENSO)之循環週期息息相關。此外，透過SSN、SOI與台灣降雨之小波相關時頻圖發現二者不僅於10至12年之頻率區間有著高度相關性，於2至8年亦呈現較高相關性，並且SSN 與台灣降雨之小波相關係數自西元1990年之後，相關係數日漸上升。
此外，本研究利用三種機器學習模型對於降雨分級預測及分類，以簡易貝氏分類器有最高之準確率，最高可達89.9%，而其中以太陽黑子對於降雨分級預測特徵貢獻度最高，所佔之比重為34%。另外，本研究改善其他類神經網路等黑盒子機器學習模型之缺點，運用貝氏網路模型分析可能影響水文水資源變動特性因子對於台灣地區雨量變動特性間之因果關係，以條件機率呈現並量化因子之間的影響程度。希冀能更了解台灣地區降雨機制並供相關政府單位更準確掌握雨量變動趨勢，進而作為台灣水資源管理方針之參考。
;In recent years, due to the impacts of climate change and Taiwan′s natural topography of high mountains and swift rivers, rainfall has become highly uneven in time and space. The dense population, along with the unfavorable development of agriculture, industry, and economy, has led to inadequate development of water resources and reduced groundwater recharge. Moreover, changes in the spatial and temporal rainfall characteristics have given rise to various hydrological and environmental problems, including extreme events. Taiwan′s main rainy seasons occur during the Meiyu season from May to June and the typhoon season from July to September. There is a significant difference in rainfall between the wet and dry seasons, but Taiwan′s water demand primarily relies on the rainfall brought by the Meiyu and typhoons. Therefore, even slight changes in rainfall patterns can directly or indirectly affect Taiwan′s risk of droughts and floods. Thus, understanding the spatiotemporal variability of rainfall in different regions and the natural factors influencing its variation, as well as predicting them, has become a critical issue of concern.
This study focuses in Taiwan , using data mining techniques to analyze observed data and remote sensing data at macro spatial and temporal scales. Wavelet signal analysis is used to extract features and explore the correlations between rainfall and factors that may affect hydrological and water resource variability. The study found that sunspots exhibit a significant 10- to 12-year cycle, with the number of sunspots varying approximately every 11 years. The Southern Oscillation Index (SOI) has a more significant correlation with a cycle of 2 to 8 years, which is closely related to the El Niño-Southern Oscillation (ENSO) phenomenon. In addition, the wavelet coherence analysis between SSN, SOI, and Taiwan′s rainfall shows a high correlation not only in the frequency range of 10 to 12 years but also in the range of 2 to 8 years. Furthermore, the wavelet coherence coefficient between sunspots (SSN) and Taiwan′s rainfall has been gradually increasing since 1990.
Moreover, this study employs three machine learning models to predict and classify rainfall levels, with the Naive Bayes classifier achieving the highest accuracy rate of up to 89.9%. It is found that the sunspot number has the highest contribution to the prediction of rainfall levels, accounting for 34% of the total contribution. Additionally, this study improves upon the limitations of other black-box machine learning models, such as neural networks, by using a Bayesian network model to analyze the causal relationships between factors that may affect hydrological resources and rainfall variation characteristics in Taiwan. The conditional probability is used to quantify the degree of influence between these factors. This research hopes to provide a better understanding of the rainfall mechanism in Taiwan and to serve as a reference for Taiwan′s water resource management policies. |
