利用邏輯迴歸與隨機森林方法建立臺灣地震型山崩臨近預測模型

Using logistic regression and random forest for nowcasting modeling of earthquake-triggered landslides in Taiwan

劉向倢(Hsiang-Chieh Liu)

指導教授 ： 莊昀叡

共同指導教授 ： 吳秉昇(Bing Sheng Wu)

國立臺灣大學/理學院/地理環境資源學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202104078

摘要

臺灣位於地震頻繁的板塊交界帶，地質年輕地形破碎，為受地震型山崩影響的高危險地區。地震型山崩是地震引發的地質災害中，對人類最具有立即性危害的災害之一，若能在地震過後迅速掌握地震型山崩的分布，可以協助減災決策。透過建立臨近預測模型，可以在地震過後迅速預測山崩分布，提供主要受災區域的資訊。統計與機器學習方法被大量用於建立崩塌潛勢模型研究，本研究透過統計與機器學習兩種方法，建立臺灣地震型山崩模型，並討論其異同。統計模型中選擇被廣泛使用的邏輯迴歸；機器學習方法中，選擇整體學習方法中常被使用的隨機森林。本研究使用1999集集地震山崩作為訓練資料，1998瑞里地震作為驗證資料，根據前人研究挑選並製作影響地震型山崩的因子，輸出的資料空間解析度為40公尺，透過山崩率的直方圖與計算點二項相關係數以及克拉瑪係數篩選地震型山崩的影響因子，再分別以邏輯迴歸與隨機森林篩選變數重要性並分別建立模型，同時測試隨機森林變數透過相關係數篩選後建立模型與隨機森林自行篩選所建立的模型預測的結果。研究結果顯示，在臺灣的地震山崩事件中，PGA以及PGA與粗糙度的共同作用，在這兩種方式建立的模型所提供的變數重要性中，相較其他因子都具有更高程度的重要性。利用此兩種方式建立的地震山崩模型都具有不錯的預測能力，根據模型的特性，當自變數眾多時，使用邏輯迴歸測試所有變數的排列組合的過程會變得相對複雜，此時可以使用隨機森林方法。若是在訓練資料相對稀少，或是自變數數量較少的情況、以及希望短臨近預測時間時，可以使用邏輯迴歸。

關鍵字

同震山崩 ； 地震引發山崩 ； 塊體運動 ； 點二項相關 ； 機器學習

並列摘要

Since the island of Taiwan is located at the plate boundary zone with frequent earthquakes, weak geology and rugged topography, it is a high-risk area affected by earthquake-triggered landslide. Earthquake-triggered landslides are one of the most immediate hazards to human beings among geological disasters caused by earthquakes. It will be helpful for hazard mitigation decisions to understand the distribution of earthquake-triggered landslides as soon as possible after an earthquake. In order to rapidly estimate landslide locations in near real-time, a nowcasting model is an effective way to provide possible landslide distributions when an earthquake occurred. In this study, I aim to assess the predictive ability of two models, logistic regression (LR) and random forest (RF). I use landslide data induced by the 1999 Chi-Chi earthquake as training data and the Jueili earthquake-triggered landslide as the validation dataset. Variables were made in 40m resolution as same as DEM. I Use point biserial correlation coefficient, Cramer’V, to evaluate the importance of factors. I then use the correlated variables to build the models. The results show that PGA and the synergistic effect between PGA and roughness are more important than other variables base on both models. Both of the two models established by LR and RF have good predictive ability. According to the characteristics of the models, when there are many independent variables, the process of using logistic regression to test the permutation and combination of all variables will become relatively complicated.. Logistic regression can be used when the training data is relatively scarce, when the number of independent variables is small, and when ones want a short prediction time.

並列關鍵字

coseismic landslide ； seismically-induced landslide ； mass movement ； point biserial correlation coefficient ； machine learning

參考文獻

Ali, M.Z., Chu, H.-J., Chen, Y.-C., Ullah, S., 2021. Machine learning in earthquake- and typhoon-triggered landslide susceptibility mapping and critical factor identification. Environ. Earth Sci. 80, 233.

Allstadt, K.E., Jibson, R.W., Thompson, E.M., Massey, C.I., Wald, D.J., Godt, J.W., Rengers, F.K., 2018. Improving Near‐Real‐Time Coseismic Landslide Models: Lessons Learned from the 2016 Kaikōura, New Zealand, EarthquakeImproving Near‐Real‐Time Coseismic Landslide Models. Bull. Seismol. Soc. Am. 108, 1649–1664.

Anooshehpoor, A., Brune, J.N., 1989. Foam rubber modeling of topographic and Dam interaction effects at Pacoima Dam. Bull. Seismol. Soc. Am. 79, 1347–1360.

Arias, A., 1970. A measure of earthquake intensity. Seismic Design for Nuclear Power Plants. Massachusetts Institute of Technology.

Ayalew, L., Yamagishi, H., 2005. The application of GIS-based logistic regression for landslide susceptibility mapping in the Kakuda-Yahiko Mountains, Central Japan. Geomorphology 65, 15–31.