臺灣位於菲律賓海板塊與歐亞板塊交界,常年受板塊運動影響,也使得臺灣地形陡峻、地質脆弱,加上近年來因極端氣候造成降雨分布不均,發生豪大雨的頻率遽增,以至於每遇颱風或豪雨時,常發生土石流、崩坍、地滑等坡地災害。因此,建立一套有效的山崩臨界降雨條件或門檻值,作為坡地安全監測與預警的參考是必要的。 本研究以模糊c均值聚類演算,將研究區域依自然與人為環境條件分為若干分區,每一個分區具有相似的環境條件,也具有相似的臨界降雨門檻。地層傾角、地層傾向與地表傾向之差值、地表坡角、土壤或岩層剪力強度、與道路距離、與河道距離為本研究在聚類分析中的聚類因子,將研究區域分為8個分區,並利用衛星影像圖進行崩塌地的判釋,將分析方法分為K-means與離群分析,比較兩者之準確度,結果為離群分析之準確度較佳,因此本研究採用離群分析做為崩塌地判釋的方法,而研究區域中第6分區之崩塌比最高,約24.918%。最後以非線性支持向量機建立每一個分區的降雨條件(降雨強度與累積降雨量)與發生山崩之預測模型,經過訓練與驗證程序,可以確認本研究所建立之分析模式與降雨促崩門檻值之正確性。
Taiwan is located at the junction of the Philippine Sea plate and Eurasian plate. It is affected by plate movement all the year round, which also causes Taiwan's steep terrain and fragile geology. In recent years, the rainfall distribution is not uniform because of extreme weather events, the occurrence frequency of heavy rain increases, so that the slope often occurred landslide when typhoon or heavy rain. Therefore, the establishment of a set of effective critical or threshold rainfall conditions of landslide is necessary. In this study, the fuzzy c-means clustering algorithm is used to classify the study area into eight clusters (zones) according to the natural environment conditions (dip of stratum, the difference of dip direction of stratum and slope surface, surface slope, shear strength of rock or soil, distance to road, distance to river). Each cluster has similar natural environmental conditions and collapse situation. Therefore, they also have similar threshold for rainfall condition. And use the Satellite imagery to judge the collapsed land, the analysis method is divided into K-means and outlier analysis, the results show that the accuracy of outlier analysis is better. Therefore, outlier analysis is used as the method of collapse interpretation in this study. The collapse ratio of the sixth zone of 24.918% is the highest in the study area. Finally, the nonlinear support vector machine technique is used to establish the prediction model to predict the precipitation threshold for landslide (maximum hourly precipitation and cumulative precipitation) for each cluster. After the training and verification processes, the correctness of the prediction model to establish the precipitation threshold for landslide can be confirmed.