Comparisons of filter, wrapper, and embedded feature selection for rockfall susceptibility prediction and mapping

The selection of influencing factors is very important for the rockfall susceptibility prediction (RSP). To improve the reliability of rockfall susceptibility prediction, three feature selection methods were used and compared to select reasonable influencing factors. The three feature selection methods are filter, wrapper, and embedded, respectively. Filter methods are represented by ReliefF and chi-square, wrapper methods are represented by genetic algorithm (GA) and binary particle swarm optimization (BPSO), and embedded methods are represented by L-norm minimization learning (LML) and recursive features elimination (RFE). Taking Meizhou City, Guangdong Province, China as the research area, 21 factors are preliminarily selected to establish a rockfall susceptibility evaluation system. The above six feature selection methods are applied to optimize the combination of factors, and the contribution of each factor in different methods is analyzed. Then, based on the optimized factor combination, the random forest (RF) model is used to predict the rockfall susceptibility. Finally, the performance of the models is evaluated. The results show that the main influencing factors of the rockfall in Meizhou City are annual average rainfall (the importance is 0.130), distance to the road (0.109), and spring kernel density (0.094). The BPSO-RF model has the best performance for all the metrics with the area under the receiver operating characteristic curve (AUC), Accuracy (ACC), Recall (REC) and F1 Score (FS) of 0.891, 0.818, 0.805 and 0.822 respectively. Compared with the initial RF model, the AUC, ACC, REC and FS of the BPSO-RF model are improved by 2.7%, 4.5%, 4.5% and 4.4%, respectively. The model performance of wrapper methods represented by GA and BPSO is significantly better than that of filter and embedded methods. It can be inferred that the wrapper method is based on feature subset search and considers the mutual information between features, which makes the method better in removing redundant features and optimizing RSP.