AutoGluon: A revolutionary framework for landslide hazard analysis

Wenwen Qi, Chong Xu, Xiwei Xu
2021 Natural Hazards Research  
Please cite this article as: Qi, W., Xu, C., Xu, X., AutoGluon: A revolutionary framework for landslide hazard analysis, Natural Hazard Research, https://doi. Abstract 11 Landslide hazard analysis is important to mitigate possible landslide hazards and ensure sustainable 12 development of society and economy. Integrating machine learning models into landslide hazard 13 analysis is a common but challenging task for researchers in general. In this work, we introduce a 14 revolutionary framework,
more » ... he Amazon's AutoGluon, a new open-source library of machine learning 15 models, to analyze landslide hazards related to the 2017 Jiuzhaigou earthquake in west China. We 16 use 11 mathematical models in the AutoGluon to perform the analysis. For each model, the coseismic 17 landslide inventory and 10 environmental and triggering factors are used as model inputs to perform 18 landslide hazard analysis. These 10 factors are altitude, slope, aspect, slope position, distance parallel 19 to the seismogenic fault, distance vertical to the seismogenic fault, distance to the epicenter, 20 lithology, distance to rivers, and distance to roads. The same number (4,834) of random points in 21 landslide and non-landslide areas are selected, 70% of which are used as training points, and the 22 remaining 30% used as validation points. It takes 47.33 seconds for data preprocessing and model 23 training for 11 machine learning models and the best result measured by Roc-AUC score is 0.94. Our 24 work shows that AutoGluon can greatly improve the efficiency of landslide hazard analysis. 25 Random Forest (Gini): AUC = 93.77% Random Forest (Entropy): AUC = 93.68% Extra Trees (Gini): AUC = 92.85% Extra Trees (Entropy): AUC = 92.87% KNN (Uniform): AUC = 91.17% KNN (Distance): AUC = 91.52% LightGBM: AUC = 93.13% LightGBM (Custom): AUC = 93.55% Catboost: AUC = 93.62% Neural Network: AUC = 92.22% Weighted ensemble model: AUC = 94.07% J o u r n a l P r e -p r o o f
doi:10.1016/j.nhres.2021.07.002 fatcat:2ukopbcbgbdx3fg663zge42squ