以空載光達資料進行台灣地區山崩型態測計之研究

Characterizing Morphologic Features of Taiwan Landslides with Airborne LiDAR Data

劉進金(Jin-King Liu)

指導教授 ： 史天元

國立交通大學/工學院/土木工程系所/博士(2013年)

https://doi.org/10.6842/NCTU.2013.00735

摘要

空載光達資料可用於各種淺層與深層山崩之探討。雖然空載光達資料為山崩研究開啟嶄新的一頁，但是由於許多新的分析方法有待開發，所以尚未成為普遍使用的工具。空載光達資料用於山崩研究主要可分為三方面：（1）山崩偵測，經由光達資料辨認山崩並且清點山崩的分佈特性與幾何特性，並可進而進行山崩潛感性分析；（2）山崩監測，經由多期光達資料分析山崩的變動或山崩的體積及其變化；（3）山崩辨認或山崩變動模型建立，經由光達資料產生之山崩型態測計參數建立預測與評估模式。本研究之目的即針對這三個山崩研究面向分別提出分析方法，並選擇台灣的案例進行實驗：（1）山崩辨識，以專家法與半自動法，進行台灣北部與南部之淺層山崩與深層山崩案例之探討；（2）山崩監測，提出個體山崩體積與全區山崩體積估算法，對高雄市案例進行個體深層山崩與區域性淺層山崩體積之估算；（3）山崩自動萃取建模，利用光達資料產生之山崩型態測計參數，以高雄市案例建立預測與評估模式。結果顯示，本研究所提出之各種光達山崩分析法均有其可行性，空載光達資料確實對於各種淺層與深層山崩之探討均有其效用。
首先，為了解空載光達技術用於山崩研究之潛力與限制，本論文首先彙整用於型態測計的重要的山崩分類與幾何特徵，進而對空載光達技術應用於台灣山崩的經驗予以審視。並分別就空載光達技術在山崩研究應用上的主要有利與不利因素加以探討。以申論空載光達資料用於山崩研究隱含之基本問題。
在淺層山崩辨識的探討上，本研究提出的方法包括（1）專家法：包括數值表面模型暈渲法與數值高程模型暈渲法；（2）半自動法：包括點雲密度指標法與物件導向分析法；與（3）自動法：本研究提出一個混合影像分割與物件導向法，並以傳統逐像元方式的分類法比較。在深層山崩辨識的探討上，本研究提出的方法包括（1）專家法；（2）半自動法：影像組織分析與物件導向分析法；（3）多期光達變動分析法：以三期光達資料檢視深層山崩之活動性。專家法分別以數值表面模型暈渲圖及數值高程模型暈渲圖作為專家判釋的依據，結果顯示淺層山崩判釋上，數值表面模型暈渲法優於數值高程模型暈渲法；相反地，深層山崩判釋上，反映林木遮蔽下之微地形的數值高程模型暈渲圖優於數值高程模型暈渲圖。半自動法先進行光達影像增揚處理 (enhancement)，產生之結果再由地質專家進一步分析判釋。光達點雲密度指標法提出四種光達點雲密度指標及其計算方法，其處理結果再與正規化數值表面模型暈渲法 (nDSM-shaded relief ) 專家判釋成果進行比較，結果顯示問題的關鍵在於點雲密度指標之參數的選擇。物件導向分析法以區域影像分割法 (area-based segmentation) 產生物件再以支持向量機法 (SVM)完成分類，分類結果分別與傳統逐像元分類法 (pixel-based method) 及專家判釋成果進行比較，結果顯示整體精度達93.4%，Kappa係數達0.817，物件導向法優於傳統逐像元分類法。至於深層山崩，利用影像組織分析與物件導向分析法進行之山崩活動性探討，本研究顯示台灣的地質與地形環境的複雜度在這方法仍有很大探討空間。在多期光達變動分析上，本研究以簡單數值高程模型差異法檢視三期光達資料，確實可以觀察到深層山崩之活動性。
在山崩體積估計的探討上，本研究提出兩種山崩體積估計方案，全區性山崩體積評估的方法包括：簡易數值高程模型相減法 (Difference of DEMs，簡稱Simple DoD Method)與個體山崩體積累計法；個體山崩體積評估的方法包括：山崩三維斷面法、平均切片法、與網格法等。本研究以小林山崩及高雄市納馬夏區之一個圖幅進行實例示範。此外，一個山崩圖幅之個別山崩可予以分割，從而產生每一個山崩的精確體積，因而可以進而探討山崩之面積A (m2)與體積V (m3)的冪次法則關係式，V = kAa。本研究納馬夏案例，獲得k = 0.099，a = 1.395，相關係數R2= 83.7%。此冪次法則可以反映不同地質、土壤、與風化侵蝕的特性。
在山崩自動萃取模式的建立上，本研究利用2005年與2009年兩期的光達資料建立一個山崩自動萃取的山崩型態測計模式。首先利用研究區的衛星影像進行自動山崩分類，其後以專家法，逐一檢視與修正山崩分佈圖。利用山崩分佈圖作為切割版，萃取發生山崩地區之光達山崩型態測計參數如坡度、地形曲率、物件高程模式(OHM)、OHM粗糙度、與地形濕度指數等。再統計產生山崩型態測計參數之區間值，由而建立一個二元多評準決策模式，此為一線性組合模式，所有落入山崩型態測計參數之區間值範圍者即為山崩。結果顯示此山崩模式之整體精度為64.9%。當進一步考慮將舊山崩區與河岸區納入，精度為64.4%，並沒有改善。當排除面積小於50m2 (即小於10m x 5m) 之山崩時，2005年與2009年之山崩預測精度分別成為76.6% 與72.5%，有明顯改善。成果顯示此山崩型態測計模式是有效的，唯型態測計參數的選擇與參數區間值的產生仍值得進一步探討。
總結而言，本研究針對空載光達資料用於山崩研究之三個面向分別提出分析方法，並選擇案例進行實驗，結果顯示本研究提出之方法可行，空載光達資料確實是山崩研究有用的工具。中央地質調查所在2009年莫拉克風災後推動全國光達測繪計畫，期間從2010年至2015年，其目的在獲取全區之一公尺解析力之光達數值表面模型(DSM ) 與數值高程模型 (DEM)、以及0.5公尺解析力之正攝影像。這可提供未來多期體積變化研究的基準，亦可提供本研究後續之應用與探討，以了解不同地文環境之適用性。此外，因為森林覆蓋下深層山崩不易被察覺，值得進一步開發半自動化的物建導向分析法予以探討。森林覆蓋區是多期變動監測之誤差主要來源，光達全光譜分析用於萃取森林底層較微弱的反射訊號，以增加森林地區數值高程模型的精度，並降低多期變動監測的不確定性，值得進一步探討。未來也可以考慮加入山崩誘因作為參數，併入本研究提出之山崩模式。

關鍵字

遙測 ； 自然災害 ； 空載光達 ； 地形測計 ； 影像增揚 ； 國家空間資訊基礎建設

並列摘要

Taiwan is located on the active collision zone between the Eurasian plate and the Philippine Sea plate. Mountains have a high relief, and rock formations are highly fractured and fragile. These physiographic settings are unfavorable to landslide susceptibility. LiDAR-derived data can be used to investigate any type of landslides including both shallow and deep-seated ones. Nevertheless, LiDAR data are not yet a common tool for landslides investigations though this technique has opened new domains of applications that still have to be developed. Applications of LiDAR in landslide investigations can be classified as: (1) Detection and characterization of landslides which include the recognition of landslides and their subsequent application in susceptibility analysis; (2) Monitoring of displacement or volume change of landslide bodies; (3) Modeling for the movement of landslides or the automatic extraction of landslides. The purposes of this research are to develop methods for understanding all these 3 aspects: (1) Landslide recognition for both shallow and deep-seated landslides with expert-based and semi-automatic approaches with cases from northern and southern Taiwan; (2) Landslide volume estimation for both shallow and deep-seated landslides with multi-temporal LiDAR data in southern Taiwan; and (3) Modeling landslide extraction with 6 geomorphometric features including slope, curvature, OHM (object height model), OHM roughness, and topographic wetness index which are derived from multi-temporal LiDAR data acquired in 2005 and 2009 in southern Taiwan.
For exploring the prospects and limitations of LiDAR Technology, the significant classification scheme and landslide features are concisely reviewed. Subsequently, both favorable and adverse factors of applying LiDAR data for landslide investigation are discussed on basis of the experiences gained so far in Taiwan. It is concluded that the awareness of the adverse factors is critical in using the LiDAR products for landslide investigations..
In the experiment of landslide detection by indices of LiDAR point-cloud density, classification results from the indices derived from the proposed four kinds of densities are verified by the result obtained by manual interpretation of the derived nDSM images. The datasets for this study are in I-Lan County after Typhoon Kalmaegi on 17 July 2008. The results show that a proper definition of the parameters for the indices is most critical for the detection of shallow landslides. Landslides recognition of the same area was also done by a pixel-based method and an object-oriented method combining area-based segmentation and a Supported Vector Machine (SVM) method. The geomorphometric features applied in the classification include Slope, OHM, and Shaded Relief which are derived from LiDAR data , as well as features of RGB, Greenness, and NDVI which are derived from concurrent images. This case shows the object-oriented SVM method is better than a pixel-based SVM method in classification accuracy and the most important features include slope and OHM. In addition, deep-seated landslide under forest can be detected in this area under expert-based shaded-relief analysis of micro-morphology.
In the experiment of landslide volume change with multi-temporal LiDAR data acquired in 2005 and 2010 in southern Taiwan, both regional approach and approach of individual landslides for volume estimation are raised. For the estimation of regional sedimentation, two methods are proposed: (1) a simple DoD method; (2) Method of Accumulating Individuals. For the estimation of each individual landslide, three methods are proposed: (1) Method of 3D Sections; (2) Method of Average Sections; and (3) Grid Method. These methods are texted with a deep-seated landslide (Hsiaolin Landslide) and with a selected map-sheet area in Namashia District of Kaohsiong City. Because the area and volume of each individual landslide in an area can be estimated, it is straightforward to model the relation between A (m2) and volume V (m3) of landslides, V = kAa. The result of the Ternbausan-One area shows that k = 0.099, a = 1.395, and R-squared coefficient of determination = 83.7%. The empirical formula reflects different physiographic conditions including geology, soils, climate and denudation processes.
In the experiment of establishing a geomorphological model for extracting landslides using multi-temporal LiDAR data of high accuracy and high resolution. Two sets of LiDAR data were acquired for before and after a heavy rainfall event. The landslides which took place from 2005 to 2009 were classified automatically by satellite images, and subsequently the landslides were interpreted and edited manually. Geomorphometric parameters including slope, curvature, OHM, OHM roughness, and topographic wetness index were then extracted using stencils of landslide polygons overlaid on respective thematic maps derived from LiDAR, DEM and DSM. The ranges of every parameter were derived from the statistics of the landslide area. Some selected non-morphometric parameters were also included in a later stage to account for all possible features of landslides, such as vegetation index and geological strength. The ranges of the parameters of landslides were optimized for the model by the statistics of the landslide area. The overall accuracy predicted by the model was 64.9%. When the buffer zones of old landslides and riverside areas were included, the overall accuracy was 64.4%, showing no improvement. When landslides smaller than 50 m2 were filtered, the overall accuracy reached 76.6% and 72.5% for 2005 and 2009, respectively. The results show that the geomorphological model proposed in this research is effective for landslide extraction.
In conclusion, the methods developed in this research for landslide detection, for multi-temporal volume change analysis, and for establishing a landslide extracting model are proved to be effective for the cases in Taiwan and for the airborne LiDAR data acquired. Generally, LiDAR data can be a good tool for landslides investigations. A national geohazard mapping program employing integrated airborne LiDAR and digital photography was launched by the Central Geological Survey after Typhoon Morakot hit southern Taiwan in 2009. The national mapping program, spanning 2010 to 2015, was dedicated to capture an entire territory of the country with airborne LiDAR and digital imagery. More datasets of multi-temporal and various physiographical settings are becoming available. Technique of OOA segmentation method for the detection of deep-seated landslides in dense forest should be developed especially for the high relief terrain of Taiwan. Other research topics include uncertainties of LiDAR analysis, the dependence of morphometric parameters on triggering events or geographical locations, and full waveform for detect the subtle reflection from the forest floor, thus to increase ground point densities of densely-vegetated area and to suppress the uncertainties of the DEM in this environment.

並列關鍵字

remote sensing ； natural hazards ； airborne LiDAR ； geomorphometry ； image enhancement ； NSDI

參考文獻

80. Lo, C. M., Lin, M. L., Dong, J. J., Chang, K. T., Chien, S. Y. & Huang, A. B. (2009) "Landslide Characterization and Zonation of Hungtsaiping Area Based on Topography, Image of Remote Sensing and PIV Technology", Journal of the Chinese Institute of Civil & Hydraulic Engineering. Vol.21, No.2. (2009/06), pp. 113-128.

14. Chang, K. T., Liu, J. K., Chang, Y. M. & Kao, C. S. (2010) "An accuracy comparison for the landslide inventory with the BPNN and SVM methods", Proc. Gi4DM 2010, 2010.

15. Chang, K. T., Liu, J. K., Wang, Z. Y. & Kao, Q. X. (2010) "A comparison of two OOA segmentation methods for the detection of rainfall-induced landslides using airborne lidar nDSM data", Proc. CACS 2010, 2010, pp.295-298.

103. Shih, T. Y. & Huang, C.M. (2006) "Airborne Lidar Point Cloud Density Indices", American Geophysical Union, Fall Meeting 2006, abstract #G53C-0919. 12/2006. 2006AGUFM. G53C0919S.

13. Chang, K. T. & Liu, J. K. (2004) "Landslide features interpreted by neural network method", Proc. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XX, Part B7, 2004, pp.574-579.