Optimizing Random Forest for Debris Flow Susceptibility Assessment: A Comparative Study of Advanced Algorithms

Abstract

Random forest (RF) method is widely used in debris flow susceptibility research. The algorithm's performance can be enhanced through the careful selection of influencing factors and the optimization of RF hyperparameters. The selected study area is Xiaojin County in Sichuan Province, recognized for its frequent occurrence of debris flows, serving as a critical area for analyzing debris flow susceptibility. 12 key influencing factors of debris flows have been identified and their correlation has been analyzed. Sparrow search algorithm (SSA), genetic algorithm (GA), whale optimization algorithm (WOA), and grey wolf optimizer (GWO) are employed for the optimization of the hyperparameters of RF, respectively. The model's performance was assessed using 5 metrics: ROC curve, confusion matrix, 10-fold cross-validation, iteration time, and convergence count. Based on the results, SSA-RF model demonstrates the highest accuracy (ACC) and area under the curve (AUC) values, with respective scores of 0.9629 and 0.98. It performs RF model by 0.06 and 0.1296, respectively. Furthermore, SSA-RF model demonstrates exceptional performance with regard to other assessment parameters. This observation suggests that RF model's performance experiences a significant enhancement following parameter optimization, thereby providing additional confirmation of the efficacy of optimization algorithms in improving RF model performance. In particular, the performance of SSA is noteworthy in this respect. The study results offer a robust scientific basis for the relevant departments in Xiaojin County and other high-risk debris flow locations to develop catastrophe mitigation and prevention strategies.