透過您的圖書館登入
IP:18.191.102.112
  • 學位論文

應用空間資料探勘技術於崩塌災害預警之研究 ─以高屏溪流域為例

Prediction of Landslide Hazards Using Spatial Data Mining – A Case Study for Gaoping River Basin

指導教授 : 徐百輝

摘要


台灣山多且高,地形起伏大,造成河短急流。每當颱風或降雨來臨時,若雨量過大,超過地表所能儲蓄之水量,則容易引發山崩、土石流及淹水等災害,往往造成人民生命財產之重大損失。若能利用現代科技及工具,在災害發生之前提出早期預警(early warning)資訊,將可有效降低災害所帶來的損失。欲進行災害預警,必須對各種災害的本質與物理特性有相當的了解,然而導致災害發生的原因有很多,例如降雨的多寡、地形、地貌、地質、土地利用等等,而每個致災因子對災害發生的影響度又會因為時間、地點、空間的不同而有所差異,甚至於每個因子之間也會相互影響,如何從這樣錯縱複雜的關係中,建立災情預警的機制,是一個值得研究的課題。台灣在歷經多次的颱洪災害後,相關單位收集了相當豐富的災害資料,其中包含有靜態的基礎空間圖資,以及動態的災害資料,他們皆具有空間性、時間性、多維性、大量性、複雜性等特點,如何從大量的空間資料中找出隱含其中的知識為本研究的課題之一。 本研究以海棠颱風於高屏溪流域的山崩災害事件做分析,首先以空間自相關檢定,探討崩塌與導致崩塌發生的相關因子的空間聚集。接著以羅吉斯迴歸、多層感知器與CHAID以不同樣本組合、崩塌緩衝區、導入空間權重概念等方式,比較預測成果的好壞。最後,藉由地形(高程、坡度、坡向)、岩性、距斷層距離...等環境與促崩因子─雨量,以及影像判釋的山崩區域面積為應變數,找出災害發生的臨界雨量條件,搭配氣象預報,達到災害預警的效果。根據實驗發現,崩塌樣本的所有相關因子都具有空間聚集性,所以以空間權重加入羅吉斯迴歸、多層感知器與CHAID都可將判釋正確率提升5%左右。另外,藉由崩塌緩衝區的設立,避免破壞面的不確定性,也可將正確率提升5%左右。在不同模式上,以多層感知器成果最佳,因多層感知器可藉由修正神經元的連結權重達到調整模型的效果,所以擬合效果最佳;再加入空間權重概念與崩塌緩衝區後,羅吉斯迴歸的正確率也可達75%。 藉由導入空間權重概念的羅吉斯迴歸,找出每個網格的臨界崩塌雨量,並以莫拉克颱風事件做驗證。從莫拉克颱風驗證結果,發現有過度判釋崩塌的問題,推測因為莫拉克颱風事件相對於海棠颱風屬極端氣候事件,所以產生偏估問題。透過本研究可以知道,想要藉由求出發生災害的臨界雨量,搭配氣象預報達到早期預警,必須考慮相關資料的空間相依性,並且以多個颱風進行訓練,讓成果更具可靠度。

並列摘要


Due to the particular geographical location and geological condition, Taiwan suffers from many natural hazards, such as typhoons, flooding, landslides, land debris, and earthquakes, which often cause series property damages and even life losses. To reduce the damages and casualty, an effective real-time system for hazard prediction and mitigation is necessary. There are a number of factors leading to hazard, and the degrees of influence change with time and geospatial. Each factors interact each other too much to construct the model of hazards. Relevant departments have collected rich data about history hazards, including static space data and dynamic disaster data which have spatiality, time, multi-dimensional, large quantities, complexity, and spatial characteristics. It is one of research topics to find knowledge hiding in a large amount of spatial data. The spatial characteristic includes spatial auto-correlation and spatial heterogeneity. Because of spatial auto-correlation, the factor of each unit is dependent. when dealing with spatial data by ignoring spatial characteristic, the result will be illogical or biased estimated. Spatial data mining is mining interesting knowledge, such as critical rainfall of landslide, from a great quantity of spatial data. The study case is about the landslide at Gaoping river basin in Typhoon Haitang, and the grids of Gaoping river basin will be counted Global Moran's I, Local Moran's I, and G statistic to confirm spatial aggregation. Because all factor have spatial aggregation, critical rainfall needs to be found by models which combine logistic regression with spatial weight matrix. After testing with logistic regression, multilayer perception, and decision tree, spatial weight matrix and the buffer zone around a landslide site can raise prediction accuracy. When examining the accuracy of critical rainfall with Typhoon Morakot, it is over-predicted because not only Typhoon Morakot is extreme climate but also rainfall type is too complex to be described by one rainfall factor. Therefore, for estimating better critical rainfalls, training samples must include data of many typhoons and the model should consider the concept of spatial auto-correlation, and then we can give early-warning by the critical rainfalls and weather forecast.

參考文獻


于嘉元,2009。土地使用變遷之空間自迴歸分析─以新店安坑地區為例,台灣大學建築與城鄉研究所碩士論文,台北市。
林恕德,2009。土壤重金屬含量與口腔癌發生率之空間相關性探討,台灣大學生物環境系統工程學研究所碩士論文,台北市。
許中立、張嘉倫,2004。地滑地降雨與地表移動之關係。坡地防災學報, 3(2)﹕ 11-2。
郭欣怡,2008。層級分析法與模糊層級分析法於山坡地生態及災害綜合指標權重分析,臺北科技大學土木與防災研究所碩士論文,台北市。
廖世傑,2006。條件式機率山崩預測模式,國立台灣大學土木工程學研究所碩士論文,台北市。

被引用紀錄


張世光(2016)。具有時空相關性類神經網路分類之研究─以崩塌地潛勢分析為例〔碩士論文,國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU201603263

延伸閱讀