二維卷積神經網路建置高解析度崩塌潛勢圖之研究

Abstract

現今極端氣候加劇，自然災害頻傳，全球逐漸重視氣候變遷衝擊調適策略。臺灣位於西太平洋颱風路徑要衝，根據中央氣象局資料顯示，臺灣近十年平均有4.1個颱風侵襲，颱風挾帶強風豪雨，易引發山區坡地嚴重的崩塌災害，進而衝擊保全對象，並導致生命財產損失及交通道路阻斷。 本研究提出一個崩塌潛勢評估模式，藉由機器學習及深度學習法建立潛在環境因子及動態誘發因子與崩塌發生之非線性關係，本模式能有效評估由降雨引起之大範圍淺層崩塌。其中，潛在環境因子可分為地形、地質和區位因子，而臺灣崩塌地主要由降雨導致的淺層崩塌，因此本研究採用累積降雨作為動態誘發因子。引起崩塌因子眾多且複雜，各因子間又具有高度非線性之關係，在參考過去文獻中，證實機器學習在處理非線性問題表現優異，故本研究在機器學習法採用近年較新穎之極限梯度提升演算法(Extreme gradient boosting, Xgboost)、多層感知機(Multilayer perceptron, MLP)及支援向量機(Support vector machine, SVM)，在深度學習法採用卷積神經網路(Convolution neural networks, CNN)二維圖像分析，由於崩塌範圍、分布及地文因子等差異性，網格解析度對崩塌潛勢模式準確度有一定的影響，因此本研究選定解析度為5 m及40 m之網格，搭配上述4種方法，並採用評鑑指標評估模式表現及其最佳化網格解析度。 本研究提出Xgboost搭配網格解析度40 m之模式(Xgboost-40m)及CNN搭配網格解析度5m之模式(CNN-2D5m)，選用宜蘭蘭陽溪流域驗證模式準確度。結果顯示Xgboost-40m模式測試指標ACC、TPR、TNR和AUC分別為78.4%、 85.7%、71.1%和0.86，表現優良且Xgboost-40m模式推估快，適用於大範圍崩塌潛勢評估；而CNN-2D5m模式測試指標ACC、TPR和TNR分別為86.3%、87.3%、 85.4%，其中AUC值也高達0.93，顯示模式對崩塌潛勢評估有良好之表現，CNN-2D5m模式建議用於高解析度需求的應用，如河道、道路等，可評估較細緻之崩塌及非崩塌潛勢值。 本研究根據不同解析度和模式所映射之崩塌潛勢地圖，能有效推估蘭陽溪流域崩塌發生區域，可供未來崩塌預警系統之參考、相關單位政策施行之依據，以達到災害預防之目的。

In recent years, extreme climate has led to many environmental problems and increased risks of natural disasters. As a result, the researchers around the world are gradually attaching importance to adaptation strategy to climate change. Taiwan is located on the main track of western Pacific typhoons. According to Central Weather Bureau in Taiwan, an average of 4.1 typhoons have hit Taiwan per year in the last decade. Typhoons accompanied by heavy rainfall are prone to causing destructive landslide disasters in the slope area. Meanwhile, landslide causes enormous losses of human life, property, and devastations to environment. Therefore, landslide susceptibility models can efficiently mitigate the disasters are desired. In this study, three novel machine learnings and convolutional neural network (CNN-2D) deep learning algorithms are employed to construct landslide susceptibility models, which can estimate the large area of shallow-seated landslide susceptibility triggered by rainfall. The three novel machine learnings include the extreme gradient boosting (Xgboost), multilayer perceptron (MLP) and support vector machine (SVM). Moreover, eleven landslide conditioning factors are used in the landslide susceptibility analysis, such as topographic factors (elevation, slope, aspect, curvature, profile curvature, plan curvature and topographic wetness index), geological factor (lithology), regional factors (distance to river and distance to road), and triggering factor (24h maximum accumulated rainfall). However, the distribution of landslides and topographic factor heterogeneity still impose challenges in selecting an appropriate spatial resolution for landslide susceptibility analysis. In order to identify the optimum spatial resolutions, the performances of above-mentioned four algorithms togethering with eleven landslide conditioning factors extracted from 5 m and 40 m digital elevation model (DEM) are checked by the accuracy and the area under the receiver operating characteristic curve (AUC). The Xgboost with 40m spatial resolution (Xgboost-40m) model and the CNN with 5m spatial resolution (CNN-2D5m) model outperform over the other models in the Lanyang river. The results in test phase show that optimized models based on Xgboost-40m (ACC = 78.4%, AUC = 0.86) and CNN-2D5m (ACC = 86.3%, AUC = 0.93) have good consistency between the correctly prediction and actual landslide susceptibility maps. By referring to the landslide susceptibility maps, Xgboost-40m which can estimate landslide susceptibility quickly is suitable for the large study area and CNN-2D5m is able to produce more accurate landslide or non-labdslide susceptibility indexs like rivers, roads and so on. Landslide susceptibility maps obtained from the proposed models are expected to be helpful to local administrations and decision makers in disaster planning.