淺層崩塌機率警戒雨量推估模式之建置 -以濁口溪流域為例

Abstract

本研究考量水文因子及土壤參數之不確定性下，發展一淺層崩塌機率警戒雨量推估模式，水文因子包括降雨特性(降雨延時、降雨量及最大降雨量)，而土壤參數包括土壤有效凝聚力、抗剪摩擦角、土壤單位重、水力傳導係數及水力擴散係數。模式建置步驟如下：(1)首先蒐集研究流域之雨量站多場降雨事件資料，並擷取降雨特性；(2)採用集群分析，將其降雨事件分類為中央型、前進型、均勻型及後退型四種雨型；(3)採用多變數蒙地卡羅法模擬衍生4,000組降雨事件與TRIGRS模式土壤參數；(4)輸入TIRGRS模式推得各場之相對應崩塌安全係數值；(5)分別以1、1.1、1.12、1.15、1.2等五個安全係數臨界值，求得安全係數小於臨界值之發生時間，並往前推算1、3、6、12、24、48、72小時等七個預警時間情況下之警戒雨量；(6)將上述所推得不同預警時間警戒雨量值與水文因子及土壤參數進行多變量迴歸分析，篩選出對警戒雨量值較高敏感度之最大降雨量、土壤有效凝聚力、抗剪摩擦角及土壤單位重等四個相關因子；(7)運用改良一階二矩法(AFOSM)求出在各假設警戒雨量值之超越機率，並將超越機率與雨型、安全係數臨界值、預警時間及警戒雨量進行羅吉斯迴歸分析，建立超越機率計算方程式。藉由所建置之崩塌機率警戒雨量推估模式，在已知預警時間及雨型情況下，可量化特定警戒雨量值之可靠度。 本研究選用濁口溪流域為例，蒐集流域內多納雨量站1967年至2013年間共1,111場降雨事件之資料作為模式建置之用。首先採用集群分析法將1,111場降雨事件區分為前進型、中央型、均勻型及後退型。另先採用濁口溪於敏督利颱風、海棠颱風及莫拉克颱風之崩塌調查資料作研究區域TRIGRS模式建置及驗證。將所建置之淺層崩塌機率警戒雨量推估模式參考水保局訂定土砂災害警戒雨量值，將其假定為崩塌警戒雨量進行可靠度分析，由分析結果可知警戒雨量值皆於24小時內，安全係數即已低於臨界值，且平均來說，警戒雨量值之超越機率(即低估風險)為10%，其可靠度達90%以上，亦即當降雨量累積至警戒值時，有90%機會安全係數值會低於臨界值。假定警戒雨量值可靠度雖甚高，卻有過於高估之風險，而無法達到預警的功效。因此，可藉由本研究所發展之淺層崩塌機率警戒雨量推估模式計算不同預警時間及雨型之警戒雨量值，並同時提供其可靠度，作為警戒雨量值修改之參考。

This study aims at developing the probabilistic rainfall threshold estimation model for shallow landslide (PRTE_SL) by taking into account the uncertainties in the hydrological factors and geographic factors. The hydrological factors serve as rainfall characteristics, including the rainfall duration, depth and storm patterns. The effective cohesion of soil, the unit weight of soil, the angle of internal friction, hydraulic conductivity and hydraulic diffusivity are regarded as the geographic factors. In this study, the framework of developing the PRTE_SL model could be grouped into seven steps: (1) collect the rainfall data of rainstorm events in the study are and extract the associated rainfall characteristics; (2) classify the rainstorm events into various types using the cluster analysis based on the storm pattern; (3) generate the rainfall characteristics, which are composed of the hyetograph, and the parameters of the landslide simulation model (i.e. TRIGRS model) by using the Multivariate Monte Carlo Simulation method; (4) calculate the safety factors by means of TRIGRS model with simulated rainstorms and the TRIGRS parameters; (5) identify the time step of the calculated safety factors less than the critical values (i.e. 1.0, 1.1, 1.12, 1.15, 1.2) (called the failure time) and then accumulate the rainfall amount (named the rainfall threshold) for different early warning time (1hr, 3hr, 6hr, 12hr, 24hr, 48hr, 72hr). The early warning time stands for the period between the current time and the failure time; (6) establish a relationship between the rainfall thresholds and the aforementioned the rainfall factors (i.e rainfall depth and maximum rainfall intensity) and TRIGRS parameters using the multivariate regression analysis. In referring to the coefficients of uncertainty factors, the rainfall thresholds are sensitive to the rainfall the maximum rainfall intensity, effective cohesion of soil, the unit weight of soil, and the angle of internal friction, which are defined as major uncertainty factors; and (7) calculate the exceedance probability of rainfall threshold rainfall by Advanced First-Order Second-Moment (AFOSM) method with the rainfall estimation equation derived at the step 7 under various early-warning times and critical safety factors; and (8) derive the relationship between the exceedance probability and corresponding the rainfall thresholds, the early-warning times and critical safety factors, named the equation of calculating the exceedance probability of rainfall thresholds. In summary, the proposed PRTE_SL model involves two equations. One is the threshold estimation equation with early-warning times, maximum rainfall intensity and TRIGRS parameters of interest for various critical safety factors; and the other is the exceedance probability calculation equation with the rainfall thresholds, critical safety factors, and early-warning times considered. As a result, the PRTE_SL is expected to provide the rainfall thresholds in association with its reliability under a specific early-warning time and the critical safety factors given. The Jhuokou river watershed is chosen as the study area, and corresponding hourly rainfall data of 1,111 rainstorm events recorded from 1967-2013 at Dona gauge are used in the model development and applicability evaluation. In advance, 1,111 rainfall events are classified into four types: advanced, center, uniform and delayed types. Then, this study sets up the TRIGRS parameters for the study area Jhuokou River watershed based on the investigation of landslide data from Typhoons Mindulle, Haitang and Morakot. After that, 4000 simulated rainstorm events and TRIGRS parameters are used for developing the PRTE_SL by means of the uncertainty method (AFOSM) and logistic regression analysis. In this study, the assumption of the rainfall thresholds based on the issued thresholds by the Soil and Water Conservation Bureau are applied to demonstrate the applicability of the proposed PRTE_SL model in the reliability assessment of the rainfall thresholds of interest. The results indicate the original announced rainfall thresholds occur within 24 hours as the safety factors are less than critical values. In average, the corresponding underestimated risk approximates 10%, and this implies that the reliability of issued safety factors reach 90%. However, the issued rainfall thresholds with high reliability might hardly achieve the goal of early warning. This is because the shallow landslide has possibly taken place before the actual rainfall amount exceeds the threshold. The proposed PRTE_SL model can provide the estimation of the rainfall threshold with the specific occurrence probability (named as the probabilistic rainfall threshold) for various critical safety factors and early-warning times. Therefore, the proposed PRTE_SL can not only quantify the reliability of the rainfall thresholds, but also provide the probabilistic rainfall thresholds referred to modify the announced thresholds.