Rockfall susceptibility mapping using XGBoost model by hybrid optimized factor screening and hyperparameter

Abstract

The accuracy of the evaluation model of rockfall susceptibility lies on reasonable conditioning factors and algorithm hyperparameters optimization. A rockfall geological database was created with 220 historical rockfalls and 220 non-rockfall cells, which was randomly divided into two datasets for model training (70%) and model testing (30%). 23 factors were selected to establish the rockfall conditioning factor database. The rockfall conditioning factors are selected by recursive feature elimination and combined with the hyperparameter optimization of grid search of machine learning-extreme gradient boosting. Thereafter, this work develops a coupling optimization model for rockfall susceptibility mapping. The results show that 9 main factors are selected by recursive feature elimination from the 23 conditioning factors, and the top ranking five factors are elevation, distance from houses, perennial average precipitation, distance from rivers, hydrogeology. After coupling optimization of factors and hyperparameters, the accuracy, precision and AUC value of the RF model are 0.7769, 0.7432, and 0.8246, respectively. Compared with the pre-optimized XGBoost model, the accuracy, precision and AUC value are improved by 0.0846, 0.0809 and 0.0616 respectively. The XGBoost model based on factor screening and hyperparameter optimization has good mapping performance.