以類神經網路評估石門水庫集水區崩塌特性

臺灣位處板塊交界及颱風路徑上，使得崩塌成為臺灣山區常見的災害，崩塌不僅威脅山區居民生命及財產安全，也影響水庫壽命及下游地區用水，造成經濟上的損失。為評估不同降雨及環境因子與崩塌特性之關聯性，本研究以臺灣北部地區石門水庫集水區為研究區域，蒐集林務局2004年到2016年之衛星判釋崩塌地圖公開資料，將崩塌地圖轉換成網格資料，計算此區域崩塌網格之年間變化，並繪製各年度的崩塌面積變遷圖。為配合崩塌面積及網格崩塌面積變遷率等崩塌資料時間單位為年，本研究以年度最大日雨量、全年總降雨事件數及全年降雨天數等因子量化降雨特性，輔以人工智慧自組織映射網路分析技術，依據拓樸結果探討所選降雨因子與崩塌特性間之非線性關係，最後再利用多項邏輯式迴歸了解崩塌特性與高程、坡度、岩層等環境因子間之關連性。崩塌地變遷分析結果顯示2004到2005年間及2009到2010年間崩塌擴張情形較為嚴重，且崩塌多發生在集水區的中部區域及西南部區域，而自2011年後集水區的中部區域崩塌已大幅減少，西南部區域仍持續崩塌。自組織映射網路拓樸結果將輸入資料分成9群，可將集水區內不同規模的崩塌依照不同的降雨型態加以分類；結果顯示集水區內的崩塌類型可分為大雨量造成的大規模崩塌、小雨量造成的大規模崩塌及大雨量造成的小規模崩塌等類型，有助於集水區的崩塌管理及防災計畫制定。此外，本研究發現石門水庫集水區崩塌與全年降雨事件數呈正相關；而崩塌與全年降雨天數則具有不同之鏈結關係，於年雨量週期的相對高峰時期，崩塌與全年降雨天數呈負相關，而於年雨量週期的相對低谷時期，崩塌則與全年降雨天數呈正相關。多項邏輯式迴歸分析結果則顯示研究區域內高程超過1500公尺之區域，崩塌潛勢與高程呈正比，高程越高崩塌潛勢越大；在坡度的部分，五級坡的崩塌潛勢較六級坡來的大；在地層的部分，汶水層、石底層、大桶山層及粗窟層等由砂岩及頁岩或硬頁岩互層所組成的地層的崩塌潛勢較水長流層來得大，而四稜砂岩雖較為穩定，但砂岩與頁岩夾層處仍具有較高之崩塌風險。綜上所述，本研究建議針對集水區內崩塌潛勢較高之區域進行監測，並依照不同降雨之特性及其重現週期進行防災之規劃。最後，為了解氣候變遷影響下集水區內降雨週期可能造成之改變，本研究亦建議持續監測降雨週期之變化。

關鍵字

石門水庫集水區；崩塌特性；降雨因子；崩塌地面積變遷率；自組織映射網路；多項邏輯式迴歸

並列摘要

Taiwan is located at the junction of tectonic plates and lies along the path of typhoons, making landslides a common disaster in its mountainous areas. Landslides threaten the lives and property of mountain residents, and affect the life of reservoirs and water use of downstream areas, leading to economic losses. This research investigates the relationship between landslide characteristics, rainfall, and environmental factors in the catchment area of the Shimen Reservoir in northern Taiwan. The publicly available satellite-interpreted landslide maps were gathered from the Forestry Bureau from 2004 to 2016 and converted into grid data to calculate the annual variations of landslide occurrences in the study area. Annual maximum daily rainfall, the total number of rainfall events in a year, and the number of rainy days in a year were quantified as rainfall characteristics. Artificial intelligence self-organizing map (SOM) network analysis techniques were employed to explore the nonlinear relationship between the selected rainfall factors and landslide characteristics. Lastly, based on the SOM map, multinomial logistic regression was applied to understand the relationship between landslide characteristics and environmental factors such as elevation, slope, and geological formations. The results of the landslide change analysis revealed more severe landslide expansion during the periods of 2004-2005 and 2009-2010. Landslides were more frequent in the central and southwestern regions of the catchment area, while the occurrence of landslides in the central region significantly decreased after 2011, but continued in the southwestern region. The results of the SOM network classified the input data into 9 clusters, enabling the categorization of landslides of different scales based on various rainfall patterns within the catchment area. Moreover, the results showed that the types of landslides in the catchment area can be classified as large-scale landslides caused by heavy rainfall, large-scale landslides caused by light rainfall, and small-scale landslides caused by heavy rainfall. This classification helps in developing landslide management and disaster prevention plans in the catchment. Furthermore, the SOM results revealed a positive correlation between landslides and the total number of rainfall events in a year. However, landslides exhibited different relationships with the number of rainy days in a year. During the relatively high peak of the annual rainfall cycle, landslides were negatively correlated with the number of rainy days in a year, whereas during the relatively low periods of the annual rainfall cycle, landslides were positively correlated with the number of rainy days in a year. The results of the multinomial logistic regression analysis showed that in areas with elevations exceeding 1500 meters, landslide potential was positively correlated with elevation, indicating that higher elevations had a greater landslide potential. Regarding slope, landslides had a higher potential on grade 5 slopes compared to grade 6 slopes. In terms of geological formations, formations composed of alternating sandstone and shale or hard shale, such as the Wenshui (Ws) Formation, Shiti (St) Formation, Tatongshan (Tt) Formation, and Tsuku (Tu) Formation, exhibited higher landslide potential compared to the Shuichanglui (Om) Formation. Although the Xileng Sandstone (Em) Formation was relatively stable, the interface between sandstone and shale still posed a higher landslide susceptibility. In conclusion, it is recommended to carry out long-term monitoring of areas with a higher risk of landslide occurrences within the catchment area and to conduct disaster prevention plans based on the characteristics of different rainfall patterns and their recurrence cycles. Lastly, to understand the influence of climate change on potential changes in rainfall patterns within the catchment area, it is recommended to monitor the potential changes in the rainfall patterns.

並列關鍵字

Catchment of Shihmen Reservoir ； Landslide characteristics ； Rainfall factors ； Landslide area change rate ； Self-organizing map (SOM) ； Multinomial logistic regression