An ensemble JCHAIDStar framework for improved landslide susceptibility prediction using Bagging and Random Subspace techniques

Abstract

This study developed a landslide susceptibility map for the mountainous area of northern Vietnam, located within the administrative boundary of the former Ha Giang Province and now under Tuyen Quang Province, to support the identification of landslide-prone areas and disaster risk reduction. The input data included a landslide inventory of 963 locations and 13 conditioning factors representing topographic, hydrological, geological, climatic, and human-related conditions. The dataset was divided into training and validation subsets using a 70:30 ratio. Four models, including Bagging-based JCHAIDStar (Bag-JCHAIDStar), Random Subspace-based JCHAIDStar (RSS-JCHAIDStar), Deep Neural Network (DNN), and Multilayer Perceptron (MLP), were developed and compared using model performance metrics and Taylor diagram analysis. The results show that Bag-JCHAIDStar achieved the best overall performance, with AUC = 0.988 for the training dataset and AUC = 0.950 for the validation dataset. On the validation dataset, this model also obtained ACC = 88.7%, Kappa = 0.760, and RMSE = 0.291, indicating good classification and generalization capability. The validation of the landslide susceptibility zoning map using the Frequency Ratio method showed that the very high susceptibility class had the highest landslide proportion, accounting for 52.11% of the total landslide pixels and reaching the highest FR value of 1.817. These results suggest that the Bag-JCHAIDStar model can provide useful support for landslide susceptibility assessment, disaster risk reduction, and sustainable land-use planning in mountainous areas.