臺灣三分之二以上的土地屬於山坡地,加上全球氣候異常之極端降雨,導致山崩災害頻傳,並伴隨著生命財產之重大損失,造成政府單位與社會相當大之負擔。因此,有效地評估降雨影響之坡地崩塌,可提供相關單位參考,期能減少災害導致的衝擊與傷害 本研究以南台灣曾文水庫及南化水庫集水區之部分區域為範圍,採用2009年、2010年及2013年共7場颱風及降雨事件前後之福衛二號衛星影像,運用基因演算自動化類神經網路技術進行影像判釋分類,以獲得各降雨事件前後之地表資訊,進而進行崩塌地之擷取。本研究利用徐昇多邊形推算研究區不同雨量測站所影響之範圍分析降雨之特性,並探討不同雨型、雨量與延時等特性與崩塌地區位及規模間之關係。 影像判釋結果顯示,不同時期衛星影像之一致性係數Kappa指標平均達0.75,具有中高程度的精確度。各雨量測站利用72小時之降雨延時,可推估出前鋒型、中鋒型、後鋒型、雙鋒型以及三鋒型等五類降雨型態。結果亦顯示,較少的累積雨量時,降雨型態之不同會影響崩塌數量及崩塌規模之大小,惟較大之累積雨量時,則無明顯之關聯性。此外,不論何種降雨型態,研究範圍之坡地崩塌大多發生在坡度介於20°~40°之間,惟前鋒及中鋒型之降雨型態引致之崩塌點位較集中在坡度20°~30°間,而其餘類型之降雨所引致之崩塌則較集中在坡度介於30°~40°間。再者,在類同之累積雨量下,前鋒及中鋒型的降雨型態所引致之研究區崩塌地區位高程遠高於其餘降雨型態。
About two-thirds of Taiwan’s total area is covered by mountains and hills. Coupled with the global climate change, rainfall-induced landslides often occur and lead to human causalities and properties loss. Therefore, the assessment of rainfall-induced landslides is indeed an important task. The study areas in this research are the Tsengwen and Nanhua Dam watershed in the southern Taiwan. The FORMOSAT satellite images before and after the years 2009-2013 (including 7 typhoons and rainfall events) were acquired and used. The Genetic Adaptive Neural Network (GANN) was implemented in the analysis techniques for the interpretation of satellite images and to obtain surface information and hazard log data. The scope of the impact of different rainfall stations in the study area was estimated using Thiessen's Polygon Method to explore the characteristics of rainfall. The relationship between the pattern, amount, and duration of rainfall and location and scale of landslide was also explored. The results of image classification show that the average value of coefficient of agreement is 0.75 at medium-high level. The rainfall patterns are classified into 5 types using 72-hour rainfall duration for each rainfall station: pre-peak, central-peak, post-peak, twin-peak, and tri-peak. The results also show that when the accumulated rainfall is small, rainfall pattern affect the number and scale of landslides. When the accumulated rainfall is large, there is no correlation between rainfall patterns and landslides. Furthermore, regardless rainfall patterns most landslide sites occur in slope between 20˚ and 40˚. Pre-peak and central-peak rainfall-induced landslides sites occur in slope between 20˚ and 30˚. The other rainfall-induced landslides sites occur in slope between 30˚ and 40˚. Moreover, in the case of the same accumulated rainfall, the elevations of landslides induced by pre-peak and central-peak are much higher than those induced by the others.