利用LiDAR DEM以物件導向分析搭配支援向量機進行潛在深層崩塌特徵自動化偵測

Abstract

近來全球氣候異常，因全球暖化導致極端氣候頻繁發生。台灣為山高、河流湍急且地質結構脆弱的地區，在極端氣候和地質脆弱兩因子的影響下，氣候變遷造成的災害已是一個不可忽視的議題。如民國98年8月的莫拉克颱風，在短時間內在南台灣降下約近一年平均降雨量，豪雨造成深層崩塌，尤其是高雄市小林村災情最為嚴重。為避免同樣的災情再次發生，判釋潛在深層崩塌特徵成為防範此類災害的首要工作，對此類災情提供事前防災工作之資訊。 政府方面，自莫拉克颱風後，委託中央地質調查所，使用航攝影像與光達之1公尺網格數值高程模型，搭配田野調查，針對台灣山區的道路和村落附近進行地質調查。以人工判釋深層崩塌發育過程中的地形特徵，如主崩崖、次崩崖、冠部裂隙、多重山脊與坡趾等，並繪製崩塌潛勢圖。該計畫成果獲得足夠的山區深層崩塌資訊和不同種類的山區圖資。 考量有效應用前述資訊與圖資，並據以開展潛在深層崩塌特徵自動化偵測研究工作，本研究使用光達之1公尺網格數值高程模型和中央地質調查所之崩塌特徵判釋成，首先以物件導向分析進行影像切割，再搭配支援向量機判釋潛在深層崩塌之特徵，並以連通分量標記法與閥值篩選分類成果，最後以剖面線分析偵測精確的崩崖位置。 本研究提出一個單以數值高程模型為輸入的自動化偵測潛在深層崩塌特徵的方法，以快速和大範圍之處理，提供含崩崖特徵的區塊目錄圖和崩崖點位圖。產出之成果可應用於崩塌邊界劃定、細部崩塌特徵偵測和崩塌目錄圖繪製。

In Taiwan, landslides are constantly triggered by earthquakes, typhoons, and heavy rains. In August, 2009, Typhoon Morakot caused severe destruction and hundreds of people’s deaths or injuries. One of the most affected areas, Siaolin village at Kaohsiung city, was buried by deep-seated landslides in Taiwan. In order to prevent the same disaster from happening again, the detection of potential deep-seated landslide features has become the primary task of preventing such disasters, and providing information before the disaster happens. Terrified by such an event, government started a program for manually detecting and mapping potential deep-seated landslides and features of potential deep-seated landslide near roads or villages. The light detection and ranging (LiDAR) digital elevation models (DEMs) with 1 m resolution, aerial photos and deep-seated landslide inventory maps were produced from the program. With these useful data of deep-seated landslides produced from the program, this study proposes a novel approach employing object-oriented analysis to segment LiDAR DEM and followed by labelling the segmentation, applying support vector machine to classify the scarp of landsides, using connected component labeling with threshold to refinement the classification result and marking the precise positions of the scarp features. This research proposes method with LiDAR DEM provided only is able to efficiently detect potential deep-seated landslide features with satisfactory results through an automatic work scheme. The result provides the marked regions which are contain features of potential deep-seated landslide and the position of the scarps. It can be applied to boundaries of potential deep-seated landslide demarcation and production of the landslide inventory map.